Short answer: It depends on who will run the system after go-live. If you will build a team, hire engineers or a freelancer. If the workflow touches several systems and nobody inside can own the build, hire an implementation partner that builds, trains and hands over. Whoever you hire, agree in writing who owns the code, prompts and test cases.
For most companies of 20 to 500 people, the right choice depends on one question: who will run the system after go-live? If the answer is a team you will build, hire engineers or a freelancer and keep the knowledge in-house.
If the workflow touches several systems and nobody internal can own the build, hire an implementation partner that builds, trains and hands over. Hire a strategy consultancy for a company-wide plan, not for the build, and an agency when you are happy to rent the result. Whoever you hire, agree in writing who owns the code, the prompts and the test cases.
We are an implementation partner ourselves, so we have a stake in this answer. The section near the end says where we fit and when we are the wrong choice.
What are your options for implementing AI?
Six kinds of provider sell AI implementation today. The labels blur, so judge each by what you get at the end: a plan, a working system you own, or a service you rent.
Strategy consultancy. Advises on where AI should go across the company, builds the business case and the roadmap. You get a plan.
AI or automation agency. Builds and often runs automations, chatbots or content workflows, usually on its own stack or a no-code platform, often on a monthly retainer. You get a running service.
Implementation partner. Builds one or more workflows inside your existing tools, trains the team, measures the result and hands over. You get a working system you own.
Freelancer. One specialist, paid by the hour or day. You get skilled hands, as far as one person reaches.
Your own team. Engineers or operations staff you employ. You get full control and the full cost.
Vendor forward-deployed engineers (FDEs). Engineers from the model or software vendor who embed in your company. OpenAI launched a separate Deployment Company in May 2026, with more than $4 billion of initial investment and about 150 FDEs and deployment specialists from its planned acquisition of the consultancy Tomoro.
There is a lot of money in this work. Sequoia’s Julien Bek estimates that for every dollar spent on software, six are spent on services. Gartner forecasts about $576 billion of AI services spending worldwide in 2026.
What is each option good and bad at?
Option | Good at · Bad at |
|---|---|
Strategy consultancy | Good: company-wide priorities, board alignment, change programs across many units · Bad: often stops at the plan; you still need someone to build |
AI or automation agency | Good: fast, cheap first automations in marketing, sales and support · Bad: often runs on its own accounts or platform; when the retainer stops, the system may stop |
Implementation partner | Good: one workflow across several systems, with approval steps, training and a measured result · Bad: costs more than a freelancer; a small firm has limited capacity |
Freelancer | Good: well-bounded tasks at a lower day rate than a firm · Bad: one person is a single point of failure; testing, documentation and cover are rarely included |
Your own team | Good: knowledge stays in the company; best for a long stream of AI work · Bad: slow to hire, expensive to keep, hard to staff with the right mix |
Vendor FDEs | Good: deep product knowledge and access to the vendor’s roadmap · Bad: ties you to one vendor; usually sized for large enterprise budgets |
How do you decide who should implement AI?
Five factors decide most cases. Score your first workflow against each before you talk to anyone.
Factor | Points toward |
|---|---|
Workflow complexity | One tool, clear rules: agency, freelancer or a feature your software already has · Several systems, exceptions, approvals: implementation partner or own team · Many units at once: strategy first |
Internal capability | No engineers: someone who builds and hands over to a named owner · Engineers with spare time: own team, perhaps with an outside reviewer |
Need for change management | Many users or a changed process: a provider who trains and measures · A handful of power users: lighter support will do |
Regulated data | Health, legal or financial data: a provider who signs a data processing agreement and can show where data goes · Public data only: fewer limits |
Budget | Under about €10,000: freelancer or existing features · One project: fixed-scope partner or agency · A steady stream of work over years: start hiring |
The budget lines are our rules of thumb, not sourced thresholds.
Internal capability is where most companies come up short. In Bitkom’s 2026 survey of 603 German companies, 85 percent of those not yet using AI said they lack the technical know-how to implement it, and 66 percent of all companies rated their staff’s AI skills as low.
Across the EU, lack of relevant expertise was the most common reason for not using AI among companies that had considered it (70.89 percent), according to Eurostat. Change management matters as much: 42 percent of German companies already using AI named a lack of employee acceptance as an obstacle.
What does each option cost?
Published rate data is thin, and none of it covers strategy consultancies or AI agencies in a way we could verify. Here is what exists, as of October 2026:
Freelancers, DACH. freelancermap’s Freelancer-Kompass 2026 IT breakdown (620 IT freelancers, part of a survey of about 5,400) gives a median of €95 an hour across IT disciplines and €100 for data and analytics. Self-reported, one platform.
Contractors, UK. IT Jobs Watch puts the median daily rate in contract jobs citing AI at £560 for the six months to October 4, 2026, across 3,314 quoted rates. The middle 80 percent ran from £393 to £779. These are advertised rates, not what firms charge.
Firms, US. The US government’s labor ceiling rate tool lists 166 AI job categories on its Multiple Award Schedule, with a median of $195.20 an hour and a middle half from $142.14 to $258.56. These are the maximum rates firms may charge the federal government, fully burdened: a reference point, not a commercial quote.
Your own team, US. The Bureau of Labor Statistics puts the median software developer wage at $135,980 in May 2025. Benefits make up 30 percent of private employers’ compensation costs, so one median developer costs roughly $194,000 a year. That is our estimate from two averages, before recruiting, tools or a second person for cover.
Outside help is not usually the biggest line. Only 21 percent of German companies using AI called external consulting a large cost item in the Bitkom survey, against 50 percent for data preparation. For the full cost picture, see what AI implementation costs in 2026.
What engagement shapes and pricing models are there?
Shape | How it is usually priced · watch for |
|---|---|
Diagnostic or discovery | Fixed fee for a few weeks · ask what you own at the end: a scored list, a baseline, a design you could give to anyone |
Fixed-scope build | Fixed price per phase · check what counts as a change request and who pays for it |
Time and materials | Hourly or daily rates · ask for a cap and weekly reporting |
Retainer or managed service | Monthly fee · check what happens to the system and the data when you stop paying |
Outcome-based fee | Share of measured savings · only works with a baseline both sides accept before work starts |
Embedded engineers | Team cost per month or year · insist on a written exit and handover date |
Most first projects work best as a short diagnostic followed by a fixed-scope build, with a retainer only for running costs you understand.
What should you ask before signing?
Ask every provider the same 12 questions, in writing, and compare the answers side by side.
Who owns the code, prompts, configuration and test cases when we finish? Get it named in the contract. Legal defaults rarely favor the client: under German copyright law, a client receives only the usage rights the contract’s purpose requires unless they are spelled out; only code written by employees goes to the employer automatically (§ 69b). In the UK, the author owns the copyright unless the author is an employee (section 11, CDPA), so a contractor’s code needs an assignment. In the US, software is not among the nine kinds of commissioned work the Copyright Office lists as possible works made for hire, so you need a written assignment.
Whose accounts does it run in? Code repository, cloud, model provider keys and logs should be in your company’s name from day one.
What does handover include? Documentation, a runbook for common failures, training for the owner, and a defined support period.
Can we swap the model? Ask where the model choice lives and what a switch costs. Models retire: Anthropic lists November 30, 2026 as the retirement date for Claude Sonnet 4.5 and promises at least 60 days’ notice. OpenAI promises at least six months for generally available models.
What baseline will you measure before building? Cycle time, hands-on time, error rate. Without it nobody can prove a result. See how to measure AI ROI.
What test set do you run before every change, and do we keep it?
Who actually does the work? Names, seniority, and what is subcontracted.
Can we call two references with a system in production for more than three months? Call them, and ask what broke.
Will you sign a data processing agreement, and which sub-processors touch our data, where? Under GDPR Article 28, a processor needs a binding contract and your authorization before it engages another processor, such as a model provider.
Where does a person approve, and how is it logged? See human-in-the-loop approval points.
What will it cost to run each month, and what happens to the price after year one?
How do we exit? Notice period, what is transferred, at what cost.
One more point for EU companies. Under the AI Act, a company that has an AI system developed and puts it into service under its own name counts as its provider (Article 3(3)). For most workflows that changes little. For a high-risk system it brings real obligations, so settle the roles before you commission one. This is general information, not legal advice.
What are the red flags?
Demo-only. A polished demo on the provider’s sample data, but no system running in production that you can see. Ask for a live one, with the client’s permission.
No baseline. A promised saving, say 40 percent faster, before anyone has measured how the work runs today.
“Agent” for everything. Gartner estimates that only about 130 of the thousands of vendors selling agentic AI are real, and says many use cases sold as agentic do not need agents. Anthropic’s engineering team advises “finding the simplest solution possible” (Building effective agents). A provider who proposes an agent before understanding your workflow is selling a label.
Lock-in. The system runs on the provider’s platform or accounts, prompts are not exported, and the contract has no exit clause.
A premium for a title. Gartner’s Mukul Saha warns: “Many providers now use ‘forward deployed’ as a label for implementation, professional services, solution engineering, or AI consulting.” Gartner expects fewer than 20 percent of FDE engagements through 2028 to turn recurring customer needs into product features. That is a warning about the delivery model, not about any one vendor. We covered Gartner’s main FDE prediction in build vs buy AI.
Where does Postmodel fit, and when are we the wrong choice?
Postmodel is an implementation partner. We build custom AI applications and workflow automation inside the tools a team already uses, with a person approving where it matters, then train the team and measure the result after go-live. We build and hand over one workflow at a time for healthcare, life sciences and consumer-health teams that already feel the cost in ops. Projects start at €10,000. If we’re the wrong fit, we say so on the fit check. Ongoing partnership is available on request.
We are the wrong choice when:
You need a company-wide AI strategy across many business units. Start with a strategy consultancy.
Your existing software already ships the feature. Turn it on first.
The task is one simple automation in one tool. A freelancer or agency will do it for less.
You have engineers with time. Build it yourselves, perhaps with an outside review.
You want someone to run the system indefinitely rather than hand it to your team.
Where to start
If you’re a COO or Head of Ops with one painful workflow, start there. Write down the first workflow you want to automate, score it against the five factors, and send the 12 questions to two or three providers of the type that fits. Compare the answers, not the slide decks. If you are still weighing a product against a custom build, read build vs buy AI. For budgets, see what AI implementation costs in 2026. Not sure which kind of provider fits your workflow? Book a fit check. If we’re the wrong choice, we’ll say so.
Frequently asked questions
What is the difference between an AI consultancy, an agency and an implementation partner?
A strategy consultancy gives you a plan for where AI should go across the company. An AI agency builds and often runs automations, so you get a service you rent. An implementation partner builds workflows inside your existing tools, trains the team and hands over a system you own.
What does outside help with AI implementation cost?
Published rate data is thin. freelancermap’s 2026 survey gives a median of €95 an hour for IT freelancers in the DACH region. In the UK, contract jobs citing AI advertised a median daily rate of £560. Only 21 percent of German companies using AI called external consulting a large cost item.
What should you ask an AI provider before you sign?
Ask who owns the code, prompts, configuration and test cases, whose accounts it runs in and what handover includes. Ask whether you can swap the model, which baseline they measure and who does the work. Ask for references in production, the data processing terms, the approval points, the running cost and the exit.
What are the red flags when you hire for AI implementation?
A polished demo with no system in production you can see. A promised saving before anyone has measured a baseline. An agent proposed for everything. Lock-in through the provider’s accounts with no exit clause. And a premium for a title such as forward deployed.
When is an implementation partner the wrong choice?
When you need a company-wide AI strategy, or your existing software already ships the feature. Also when the task is one simple automation in one tool, when you have engineers with time, or when you want someone to run the system indefinitely.
Related reading
Sources
OpenAI, OpenAI launches the OpenAI Deployment Company to help businesses build around intelligence, May 11, 2026.
Julien Bek, Sequoia Capital, Services: The New Software, March 5, 2026.
Gartner, Gartner Forecasts Worldwide AI Spending to Grow 49.5% in 2026, September 16, 2026.
Bitkom, Erstmals nutzt die Mehrheit der Unternehmen KI, September 14, 2026. Telephone survey of 603 companies with 20 or more employees in Germany.
Eurostat, Use of artificial intelligence in enterprises, data extracted December 2025.
freelancermap, IT-Freelancer Stundensatz 2026, July 2, 2026, based on the Freelancer-Kompass 2026.
IT Jobs Watch, AI contract jobs and rates, UK, six months to October 4, 2026.
U.S. General Services Administration, Labor Category Ceiling Rates, search “artificial intelligence”, data of October 4, 2026.
U.S. Bureau of Labor Statistics, Occupational Outlook Handbook: Software Developers, Quality Assurance Analysts, and Testers, accessed October 4, 2026.
U.S. Bureau of Labor Statistics, Compensation costs for private industry workers averaged $46.89 per hour worked in June 2026, September 17, 2026.
Gesetz über Urheberrecht und verwandte Schutzrechte (UrhG), § 31 and § 69b.
U.S. Copyright Office, Circular 30: Works Made for Hire, revised August 2024.
Anthropic, Model deprecations, accessed October 4, 2026.
OpenAI, Deprecations, accessed October 4, 2026.
Regulation (EU) 2016/679 (General Data Protection Regulation), Article 28.
Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 3(3).
Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, June 25, 2025.
Anthropic, Building effective agents, December 19, 2024.
Gartner, Gartner Predicts 70% of Enterprises Will Abandon Agentic AI Built by Vendor Forward-Deployed Engineering by 2028, September 29, 2026.