All posts ROI & Business Case

What Enterprise AI Implementation Really Costs (and What Drives the Number)

The short answer

An AI implementation consultant is priced by the day or as a fixed project fee, and the cost guides in circulation quote whole project budgets instead: one dev-shop guide puts a startup’s agentic AI build at $15,000 to $120,000 and an enterprise one at $200,000 to $1M+. The consultant’s fee is one of six cost lines, and not the one AI users name most often: in Bitkom’s survey of 603 German companies with 20 or more employees, published in September 2026, the companies already using AI named external consulting as a large cost item far less often than infrastructure, data preparation and integration into existing systems, work that may itself be bought from a consultant.

The spread is wide because the guides price two buyers in one table. A small company automating one department inside the SaaS tools it already runs is one buyer. A EUR 250M to 2B company whose workflow crosses an ERP, a CRM, a document store, an identity provider and a works council is another, and most of its number sits in lines a quote leaves off: integration against legacy systems, data readiness, the compliance layer, the implementation labour itself, change management, and the run costs after go-live, infrastructure included.

Nucleo publishes no engagement figures. What follows is the buyer’s side of the desk: which line applies to you, who leaves it off the quote, and how to price it before signing.

Key takeaways

  • The consultant’s fee is one line of six. German AI users name infrastructure, data preparation and integration as large costs far more often than external consulting, though that work may itself be bought from a consultant. Tokens come last.
  • Two buyers, one table. Published ranges mix a single-department build on existing SaaS with a rollout across several systems. Only the second describes a mid-cap workflow that crosses systems.
  • The multiple sits between pilot and production. One practitioner has watched AI-adjacent spend reach 4 to 8 times the pilot budget within 100 days of approval. Gartner says a production system can cost orders of magnitude more than the pilot.
  • After the sandbox, the number often has no single owner. In one 2026 survey, 26% of enterprises share ownership of AI economics across functions and 12% have none. Model invoices show a total, not the workflow or the team behind it.
  • Every line can be priced before signing, if the quote names the systems, the data owners, the risk class of the use, the change owner and the monthly volume.

How much does an AI implementation consultant cost?

Published rate cards differ widely by provider and role (the ones we found), and the rate rarely decides the budget. What decides it is which of the six lines the quote covers.

Bitkom, the German digital association, asked companies that already use AI which items they count as large costs. Infrastructure such as cloud and compute was named most often, data preparation almost as often and integration into existing systems third. Subscriptions and licences for AI software came well behind, level with spending on external consulting. Staff training came lower still, and token consumption last. Bitkom’s president, Ralf Wintergerst, put the finding in one line: “Wer nur eine Lizenz kauft, hat KI eingekauft, aber noch nichts verändert” (whoever only buys a licence has bought AI, but changed nothing yet).

A consultant who prices only their own days has priced one line of six. Infrastructure, data and integration land on your side of the contract unless the quote names them.

If the work is configuring a product, an implementation consultant is enough; if it is production code against a problem nobody has specified, it is forward-deployed work, and the labour line moves with it.

How much does it cost to implement AI into a business, and which business is the estimate for?

Enterprise AI development cost depends first on which of two buyers you are, and the guides rarely say. The biz4group guide draws the line itself: “Small businesses generally deploy AI within one department using existing SaaS tools, whereas enterprises invest in organization-wide rollouts that require governance, security, custom integrations, and change management across multiple teams.” Then it prices both in one article, under a headline range of $15K to $150K that stops below its own enterprise figures.

The single-department buildThe multi-system mid-cap deployment
Typical buyerA small company, or one team, automating work inside the SaaS tools it already runsA EUR 250M to 2B company whose workflow crosses departments and vendors
Systems touchedOne or two, through the vendors’ own connectorsAn ERP, a CRM, a document store, an identity provider, often one system with no modern API
Who integratesThe tool’s connectors, configured by the team or a freelancerEngineers who write and maintain integrations against each system’s permitted interface
Compliance exposureGDPR on the data the tool touches; rarely a high-risk use under the EU AI ActGDPR with an impact assessment where the risk is high, AI Act classification per use, a decision record, and in Germany the works council
Who runs itThe team that built it, on the vendor’s planA named owner per workflow, IT operations, and whoever keeps the context current
Price you can seePlan prices if a small company builds it, such as n8n’s Business plan (self-hosted, for companies with fewer than 100 employees) at 667€ a month billed annually, as of October 2026None until the systems are counted
Which guide figure appliesIf a supplier builds it, the low ranges, such as $15,000 to $120,000 for a startup buildThe enterprise ranges, $200,000 to $1M+ in the same guide, and only as a sign of scale

The second column is where money leaks. Freshworks surveyed 12,021 IT professionals, more than 9,000 of them in mid-market organisations of up to 5,000 employees, and found that “an average 25% of mid-market AI spend is lost to complexity overhead”. Only 15% have AI integrated across core business operations, 36% remain stuck in pilots, and 55% say deployment alone takes between 6 and 12 months (Freshworks). It is vendor research, and it points the same way as Bitkom’s ranking.

We found no European survey that publishes a project-cost range for a multi-system deployment at this scale.

What are the hidden costs of enterprise AI implementation?

Six lines, and the model is rarely the largest.

DriverWhat moves itWho tends to leave it off the quoteHow to price it before signing
Integration against legacy systemsSystems read and systems written; the authentication model; whether a usable API exists; whether the vendor’s API terms allow it; each system’s upgrade cadenceThe tool vendor, whose quote stops at its connectors; the integrator, when nobody counted the systemsA line per system: read or write, interface, identity, owner, upgrade calendar. A fixed price per integration with named exclusions
Data readinessWhere each answer the workflow needs lives, whether two systems disagree about it, how current it isMost suppliers, because nobody has looked at the data when the quote is writtenA data inventory for one workflow before the build, with cleanup priced as its own line
The compliance layerThe AI Act risk class of the use; whether GDPR requires an impact assessment; the decision record and its retention; human review timeDev shops and SaaS vendors; consultancies price the assessment but not the engineeringClassify the use first. Price the decision record as a deliverable and review time as a run cost
The labour itselfHow much of the problem is unspecified; configuration or production code; who carries the overrun; your own system owners’, subject experts’ and security reviewers’ time during the buildMost quotes leave off your own people’s time; a day-rate quote also leaves the total openA fixed fee per scoped workflow, or a day-rate total with a cap and named deliverables
Change managementHow many people change how they work; the works council; training; the productivity dipThe build supplier, because the people who change are yoursA named change owner and a person-day estimate next to the implementation estimate
Infrastructure and run costs after go-liveHosting and compute; vector stores, observability, guardrail and evaluation tooling; egress; model usage and vendor units; keeping the context engine current; review timeThe build supplier, whose quote ends at go-live, and every metered vendor, whose unit price can changeA monthly cost per workflow at expected volume, infrastructure included, with every metered provider’s price-change terms

Read the third column as a checklist, not a verdict on any supplier. The rows are work, not invoices: integration, data and compliance are paid for in someone’s days, the supplier’s or your own. For each row, ask whose days it takes and whether the quote counts them.

The lines furthest from the model surprise most. A 2025 survey of 372 enterprises by Benchmarkit with Mavvrik, a cost-management vendor, named data platforms the top source of unexpected AI spend (56%), with LLMs ranked 5th.

How much does integrating AI with legacy systems cost?

As much as the number of systems the workflow reads and writes, multiplied by how hard each one is to reach. No guide can give you that figure. What you can get before signing is the count.

Integration is where mid-market programmes stall: in the Freshworks survey, system integration complexity (27%) is the top reason pilots fail to become full programmes, narrowly ahead of skills shortages and configuration. The surface is growing. In Salesforce’s 2026 connectivity benchmark, the average number of applications in use at German enterprises of 1,000 or more employees rose from 844 to 920 in a year, and fewer than a quarter of them are integrated with each other (Salesforce Germany). That figure rests on a German subsample of 100 IT leaders, in research by a vendor that sells integration.

Five things move the integration line:

  • Whether the agent reads or writes. Reading an order from the ERP needs a query; posting a credit note needs the business rules, the permissions and a way to reverse a mistake.
  • The identity it acts under. Acting under each user’s own permissions through single sign-on respects the access model you already govern; a shared service account is quicker to build and becomes a second access model to audit.
  • The interface: a documented API, a file export or a screen. ECC customisations and mainframes are the expensive end: see the Z-code in an ECC system and why mainframes are the hardest legacy problem.
  • Permission to connect. What software outside SAP may do with SAP’s APIs is now a contract question, and the answer can add a licence line.
  • Upgrades. The integration surface moves every time one of those systems is upgraded, which is why the cost of a build lands in year two.

To price it, ask for those five points as a per-system schedule, with the owner on your side named for each system. A quote that prices “integration” as one line has not done the count.

How much does data preparation add to an AI project?

Compiled guides quote it as a share of budget, but no survey we found measures it. Bitkom’s AI users name it as a large cost almost as often as infrastructure, and more often than integration, licences or consulting. In a survey of 600 data leaders by Informatica, a data-management vendor, “Even as companies race to production, 57% admit data reliability is a top barrier” (Informatica).

Four questions move the line, asked for every decision the agent makes. Where does the answer live: an ERP field, a PDF, an email thread? Do two systems disagree about it, as sales and finance do about what a customer is? How current is it? Who keeps it current after go-live?

The expensive case is the answer that lives in no system at all: the exception a planner handles from memory, or the reason a claim was accepted last time. Writing that down is the work of building a context engine, the owned layer of definitions, procedures and past decisions that agents read from. What building a context layer involves sets out the steps.

What does EU AI Act and GDPR compliance add to the cost of an AI agent?

Work in four places: classifying each use under the AI Act, a data protection impact assessment where GDPR requires one, a decision record and its retention, and human review time that recurs every month. Most quotes price none of the four as engineering.

German companies expect the work to be heavy. Of those that are affected by the AI Act or are checking whether they are, almost nine in ten expect a high implementation effort and about half a very high one, per Bitkom’s survey.

The European Commission’s own per-system estimate comes from its 2021 proposal. It put compliance costs at approximately EUR 6000 to EUR 7000 for the supply of an average high-risk AI system of around EUR 170000 by 2025, verification costs of another EUR 3000 to EUR 7500 for suppliers of high-risk AI, and, for users, the cost of human oversight at approximately EUR 5000 to EUR 8000 per year (COM(2021) 206). Those figures predate the final text and the 2026 amendments, and cover only high-risk systems. Treat them as the regulator’s 2021 assumption, not as your budget.

The four pieces of work:

  1. Classification. Agents for work such as invoice matching or service triage are usually not high-risk. Uses in recruitment, worker management or the creditworthiness of individuals can be, and German companies affected by the AI Act expect on average more than two high-risk systems in use, per Bitkom, so classify each use rather than assume. The classification decides whether the rest of the compliance budget is small or large. The obligations and their dates are in what the EU AI Act demands, and from when. The AI Omnibus extends some SME measures to small mid-cap companies, a defined EU category based on headcount, turnover and balance sheet total, not a synonym for mid-cap. Check whether you fall inside it before a quote assumes the lighter regime.
  2. The impact assessment. GDPR requires one where processing is likely to result in a high risk to people, and an agent reading HR or customer records can meet that test. The controller carries it out, so ask the supplier to document the data flows the assessment needs.
  3. The decision record. What the agent saw, which rule it applied, who approved the outcome. It is engineering, it belongs in the build quote, and audit trails for AI agents covers what it has to hold.
  4. Review time. Someone looks at the cases the agent routes to a person, every month the system runs. That is a run cost priced at your own people’s time, and graduated autonomy is how to size it to the stakes.

How much does change management add to an AI implementation?

In one practitioner’s rule of thumb, change management and business disruption together take as much effort as the implementation and up to twice as much, with training on top. Even among German companies already using AI, Bitkom’s survey finds a lack of acceptance among staff named as an obstacle, alongside data protection and legal uncertainty.

Rohit Saini, vice president for data and AI growth at EXL, a data and AI services firm, draws the rule from large-scale enterprise rollouts: “For every 100 person-days of implementation effort, expect roughly 25 person-days of training effort and a further 100 to 200 person-days of change management and business disruption costs” (FP&A Trends). It is one person’s observation, not a survey. He adds that CFOs who benchmark AI investment purely against the build effort consistently underestimate total programme cost by a factor of 2 to 3.

In Germany a system that records who approved what can fall under the works council’s co-determination right (what that means in practice); count the weeks that agreement takes in the change estimate.

Why do AI costs multiply between pilot and production?

Because the pilot did not pay for most of what production needs.

Saini has observed in enterprise rollouts that, 100 days after a pilot is approved, “total AI-adjacent spend has grown 4 to 8 times the original pilot budget, and yet nobody can fully explain where the money went”. Gartner, quoted by THE Journal, says “Creating a production-ready GenAI system can be orders of magnitude more expensive than running a pilot,” and expects that through 2028 at least 50% of GenAI projects will overrun their budgeted costs, because of poor architectural choices and a lack of operational know-how. The 3 to 8x multiple that circulates in budgeting guides appears in compiled planning ranges such as iternal.ai’s 2026 guide, which calls its planning ranges directional, not quotes, and cites another planning guide in support of the multiple. We found no primary study behind it.

The money goes to four places.

  • Infrastructure sized for a demo. Production adds vector stores, monitoring and guardrails, priced under run costs below.
  • Integration found only on real systems. A pilot that ran on a sandbox copy has not met the identity provider, the rate limits or the record two systems disagree about.
  • Change management the pilot never did. A pilot’s users are volunteers; production’s are everyone.
  • Volume. A contributed column on CIO.com describes the pattern: a company approves an agent because it works in the pilot, then reconsiders once thousands of people use it and every task triggers several paid requests.

To keep the multiple small, make the pilot meet production’s problems early: run it on the real systems and identity provider, and price production before the pilot starts. The four gates from pilot to production set out what each stage has to prove and who owns it.

Who owns AI spend once a pilot goes live?

Most often a technology function, sometimes several functions jointly, and in 12% of enterprises nobody, per the Tokenomics Foundation survey below. We found no survey that measures ownership per workflow. The pilot had a sponsor and a budget line. Production has a model invoice, a cloud bill, several vendor meters and a team that has moved on.

Budgets are approved once a year. Agent spend compounds with every call: more users, more cases, more steps per case. The same CIO.com column, by Carter Page, an executive at the software vendor Astronomer, describes the result: “Companies know how much they spent on AI, but they often cannot explain which workflows generated those costs, whether that spending created meaningful business value, or which teams should ultimately own it.”

In the FinOps Foundation’s State of FinOps 2026, allocating AI costs to business units is among the top AI challenges practitioners cite. The Tokenomics Foundation asked 472 enterprises, with a median revenue of $1.8B, who owns AI economics. In its State of Tokenomics, a release candidate published in September 2026, 88% have defined ownership and 35% name a CTO, CIO or technology function, but 26% say ownership is shared across functions and 12% have none. Of that last group, the survey finds: “Not one of them can connect AI spend to a CFO outcome”. Those with defined ownership were 3.7x more likely to show value to the CFO.

Settle ownership before production, per workflow, not per programme:

  1. One owner per workflow, accountable for the number the workflow should move and for what it costs to run. The FinOps Foundation’s guidance on AI spend gives the reason: “AI spend sits between Finance, Engineering, Security, and Procurement, so it needs a named, accountable owner.”
  2. Every model call and vendor unit tagged with the workflow that caused it.
  3. A monthly cost per case, reviewed against the approved baseline.
  4. A stopping rule, agreed before go-live: the cost per case, or the result, at which the workflow is paused and reviewed.

A build handed over at go-live leaves the supplier no stake in what the system costs or achieves. A run fee partly tied to an attributable outcome gives it one; ask separately who pays the model and vendor meters, because that decides who watches the cost per case.

Why is AI so expensive to run after go-live?

Because every case costs something, and the meters multiply. For a company that calls models through an API, most of the model cost is usage, billed per token or per unit unless it buys provisioned throughput or a committed spend, and it recurs every month. Bitkom’s AI users still name tokens last among their large cost items. What grows after go-live is the volume of cases, multiplied by the calls each one takes.

The run cost has five parts.

  • Infrastructure: cloud and compute, the cost Bitkom’s AI users name most often, plus egress, vector stores and the observability, guardrail and evaluation tooling. Saini says the vector stores and tooling can, taken together, exceed the model API bill.
  • Model usage, which follows cases and the steps in each, not licensed seats: an agent can trigger several model calls, retrievals and tool calls for one task. Provider prices also change on the provider’s schedule: in October 2026, Google’s Gemini API pricing page listed paid-tier input and output prices for Gemini 4 Argon doubling from 1 December 2026.
  • Metered vendor units. SAP bills agents in AI Units per action, Salesforce in Flex Credits, and Palantir converts tokens into compute-seconds. Each unit has a price; what a case costs depends on how many units it takes.
  • Keeping the context engine current. Definitions change and records age. Whether your people or a supplier on a run fee keeps them current, the cost is yours and needs an owner on your side: the bench problem in another form.
  • Review time, priced at the hours of the people who look at the cases the agent routes to them.

Bitkom’s AI users name hard-to-calculate costs among their main obstacles, behind data protection and legal uncertainty. The Benchmarkit and Mavvrik survey found that “80% of enterprises miss their AI infrastructure forecasts by more than 25%”. The fix is arithmetic you can do before signing: per workflow, a monthly volume V times the model calls or vendor units per case U times the unit price P, plus infrastructure and review time. The figure will be wrong, but it gives the owner a baseline to be wrong against.

How do you price an AI implementation before you sign?

Agentic AI implementation cost can be priced line by line, against what the process costs you today. A quote that arrives as one number has either counted everything or nothing, and you can tell which by asking for the baseline and the six lines.

  1. A baseline: what the workflow costs today in hours, errors and delay, measured before anyone builds. The ROI formula is the value side of the same sum; this post is the cost side.
  2. A per-system integration schedule, with a fixed price per integration and the exclusions named.
  3. A data inventory, the cleanup priced with acceptance tests, and an owner per source.
  4. A risk class for each use, the impact assessment if needed, the decision record as a deliverable with a stated retention period, and review hours as a run line.
  5. Labour as a fixed fee per scoped workflow, or a capped day-rate total with named deliverables and a statement of who pays for an overrun, because a day-rate contract sells effort rather than a working system. Count your own people’s days next to the supplier’s.
  6. Named owners for training and change, a person-day estimate next to the implementation estimate, the productivity-dip assumption and the works council’s timeline.
  7. A monthly run cost per workflow at expected volume, infrastructure included, the price-change terms of every metered provider, and the owner of the number after go-live.

Vendor ROI calculators will not do this for you: Salesforce’s own says implementation costs are not included. The contract questions that sit beside the price, such as who owns the result and who maintains it in month thirteen, are in what to settle before you sign and the questions to ask a partner before you sign.

We price the way we ask buyers to read a quote: a fixed fee to build, scoped against what the process costs you today, a recurring fee to run and evolve it, and part of the fee tied to the outcome where it is a discrete event we can attribute (how that is structured). There is no rate card, so we take longer to compare than a supplier who has one.

When is the cheaper AI implementation the right one?

When the workflow lives inside one product and touches one system. Then the vendor’s packaged agent or a no-code build is cheaper, faster to start and easier to run, and an engagement like ours is the wrong purchase.

Take the cheap number when:

  • the work starts and ends inside one suite, such as a period-end close in S/4HANA Cloud, where the vendor’s agent already has the semantics, the permissions and the audit trail (when SAP Joule is enough on its own);
  • a small team automates its own work in tools it already uses, at published plan prices such as n8n’s, and the people who build it will run it;
  • the use is not high-risk and touches no personal data beyond what the tool already holds;
  • there is no number to move yet, so the first job is finding the workflow worth automating, not pricing a build.

The expensive mistake is the opposite: paying the first column’s price for the second column’s work, and discovering integration, data and compliance after the pilot.

FAQ

How much does an AI implementation consultant cost? An AI implementation consultant is priced by the day or as a fixed project fee, and the cost guides in circulation quote whole project budgets instead: one dev-shop guide puts a startup’s agentic AI build at $15,000 to $120,000 and an enterprise one at $200,000 to $1M+. The fee is one of six cost lines. In Bitkom’s 2026 survey of German companies with 20 or more employees, AI users named infrastructure, data preparation and integration as large cost items far more often than external consulting.

How much does it cost to implement AI in a business? It depends on which business. A small company automating one department in SaaS tools it already runs pays plan prices and some configuration time. A EUR 250M to 2B company whose workflow crosses an ERP, a CRM, a document store and an identity provider pays mostly for integration, data, compliance, change and run costs, which have no price until the systems are counted.

What are the hidden costs of AI implementation? Six lines that quotes tend to leave out or leave open: integration against legacy systems, data readiness, the compliance layer (AI Act classification, a GDPR impact assessment, a decision record, review time), the labour itself, whose total a day-rate quote leaves open, change management, and infrastructure and run costs after go-live, including keeping the context agents read current. Most quotes also leave off your own people’s time during the build.

Why do AI costs increase from pilot to production? The pilot did not pay for most of what production needs: infrastructure beyond the model, integration against real systems, change management for everyone, and volume. One practitioner writing in FP&A Trends has observed AI-adjacent spend reach 4 to 8 times the original pilot budget within 100 days of the pilot’s approval. Gartner says a production-ready GenAI system can be orders of magnitude more expensive than the pilot.

Who should own AI spend after go-live? One named owner per workflow, accountable for the number it should move and for what it costs to run. In the Tokenomics Foundation’s 2026 survey, 12% of enterprises had no owner for AI economics, and those with a defined owner were 3.7x more likely to show value to the CFO.

Why is AI so expensive to run? Every case costs something, and the meters multiply: hosting, vector stores and monitoring, several model calls per agent task, vendor units such as SAP AI Units or Salesforce Flex Credits, human review time, and keeping the context agents read current. For a company calling models through an API, most of the model cost is usage, billed per token or per unit unless it buys provisioned throughput or a committed spend, and it grows with the number of cases and the steps per case, not with the number of licensed seats.

How do you estimate agentic AI implementation cost before signing? Line by line, against what the process costs today: a per-system integration schedule, a data inventory, a compliance classification, labour on a fixed fee or a capped total, a change estimate with named owners, and a monthly run cost per workflow at expected volume, infrastructure included. A quote that arrives as one number has not counted the lines.

When is a cheaper AI implementation the better choice? When the workflow starts and ends inside one product and touches one system, or a small team automates its own work in tools it already uses. Then a packaged agent or a no-code build is cheaper and easier to run, and a forward-deployed engagement is the wrong purchase.



Nucleo is an AI-native transformation company for European mid-caps. Our forward-deployed engineers build the context engine your company runs on, inside your stack and model-agnostic, then the systems that do the work on top of it, priced against what the process costs you today. See how an engagement works, or write to us.

Sources: biz4group; Bitkom, September 2026 survey; n8n pricing; Freshworks; Benchmarkit and Mavvrik, via PR Newswire; Salesforce Germany, Connectivity Benchmark 2026; Informatica; COM(2021) 206, archived; European Commission, AI Omnibus; European Commission, SME and SMC definition; EU AI Act, Annex III; GDPR, Article 35; FP&A Trends; THE Journal, on Gartner; iternal.ai; CIO.com; FinOps Foundation, State of FinOps 2026; Tokenomics Foundation, State of Tokenomics; FinOps Foundation, FinOps for AI; Gemini API pricing.

The future belongs to those who see it before it is obvious, and build it while everyone else is still out there scouting for the best AI tools.