Who sells AI implementation in 2026?
The AI labs and clouds sell it through their own deployment arms. Large integrators sell it too, and so do boutique studios and freelancers.
The forward-deployed label now shows up in every one of those groups, so it tells you little about who will sit with your team.
Buying outside help is the normal route for mid-market companies now. In RSM's July 2026 survey of 1,030 middle-market companies in the US and Canada, all already using AI, 52% use outside consultants and 58% plan to spend $1M or more on AI in 2026. Among those with only moderate or limited pilot success, the top blockers were data quality (53%) and integration (47%). Both are implementation problems, which is why the partner you pick matters more than the model you pick.
Each kind of seller fits a different buyer.
| Seller | Examples in 2026 | Fits | Watch for |
|---|---|---|---|
| Vendor deployment arms | OpenAI Deployment Company (more than $4B raised), AWS ($1B FDE organization), Microsoft Frontier Company ($2.5B, 6,000 experts), Ode (about $1.5B), Salesforce (a 1,000-person FDE team) | Companies already committed to that vendor's models or cloud. Ode is aimed at mid-size companies, starting with its backers' portfolio companies. | The partner is also the vendor, so model and cloud choices aren't neutral. Gartner's 2028 warning is about this group. |
| Large integrators | Accenture, Deloitte, SHI | Enterprises that need many teams in many places under one contract. | FDE washing. Ask who will sit with your team and what they have put into production. |
| Boutique studios | DK Studio and firms like it | Mid-market companies that want one workflow in production without a long program. | A thin bench. Ask who covers when the lead engineer is away. |
| Freelancers | Independent FDEs and fractional AI leads | A scoped piece of work, like an architecture review or a stalled integration. | One person and no backup. The handover depends on their notes. |
Price follows the same split. Gartner puts vendor FDE consulting fees at up to about $200,000 per quarter per use case. Small AI studios quote a paid diagnostic at around $5,000-$15,000 and one workflow to production at around $25,000-$50,000, and fractional AI-lead contracts in New York ran $175-$300 an hour in the listings we tracked.
What Gartner's FDE warning means for buyers
Gartner expects 70% of enterprises to abandon agentic AI built by vendor FDEs by 2028, because costs climb and the client can't evolve the system on its own. For a buyer, that moves ownership and the exit plan to the top of the negotiation, ahead of the model.
The same research says fewer than 20% of FDE engagements will feed the vendor's core product. If most of what you pay for never improves the product, it is custom work, and custom work should come with your name on the code.
Gartner also warns about "FDE washing", its term for ordinary consulting relabelled as FDE. Some sellers say as much themselves. In a report by The Register, the integrator SHI and the vendors AWS and Cisco said embedding engineers at customers is not new, and DoIT moved its long-running customer reliability engineering practice under the FDE name in late 2025.
Gartner's advice is to settle scope, incentives, governance, ownership, knowledge transfer and the exit plan from day one. Knowledge transfer looks like the step buyers skip. BCG's January 2026 survey of 100 PE investors found that 70% of successful firms use specialized digital boutiques, and that only 45% make sure knowledge transfers to their own teams.
AWS, at least, puts the right goal in public: it says it built its $1B FDE organization so customers are self-sufficient when a deployment ends. Whoever you hire, ask them to write the same goal into your contract.
12 questions to ask an AI implementation partner
Send these before the first meeting and ask for written answers. If the first two come back vague, you can skip the rest.
1. Who exactly will do the work?
Get names. The senior person in the pitch should be the person in your Slack, and you should know how many days a week they spend with you. Ask about on-site time too: OpenAI's own FDE postings ask for up to 50% travel, and Palantir's FDEs spend around 25% of their time on site, according to Pragmatic Engineer.
2. What do you have running in production right now?
Demos prove a model can do a task once. Ask for a system people use every day, and ask what broke in it last month. Teams with real production work answer the second part quickly, because they remember.
3. How do you decide where AI belongs?
A good partner goes step by step. Some steps need a model, others need plain code, and anything that moves money needs a human approval. If every box on the diagram says "agent", you are paying for agents you don't need. Our accounts payable breakdown shows that sorting for one workflow.
4. How will we know it works?
Ask for the eval plan before the build: a test set made from your own past cases and a pass mark agreed before go-live. A partner who plans to "tune it after launch" is planning to test in production. Missing evals are one of the reasons AI pilots stall before production.
5. Who owns the code and the IP?
You should, in writing. The code belongs in your repository, prompts and evals included, with the IP assigned to you when you pay. If the system only runs inside the partner's platform, the exit costs whatever they decide.
6. What is the exit plan, and how does knowledge transfer?
Ask what your team will be able to change on its own after the partner leaves, and how they will learn it. Good answers include a written handover and time spent pairing with your engineers. A partner who can't name the point where you stop needing them hasn't planned an exit.
7. Where does our data go?
Ask which model provider processes it, in which region, how long it is kept and whether anyone trains on it. Then ask the same about every other tool in the chain, including the observability tools that store prompts and outputs.
8. When does the security review happen?
At the start. A security review booked for the week before launch is how a finished pilot ends up waiting on a questionnaire. Put your IT or security lead in the first meeting and hand the partner your questionnaire before anyone writes code.
9. How do you price, and what is fixed?
Fixed prices per workflow, monthly embedded capacity, hourly contracts and per-quarter vendor fees all exist in 2026. Ask what is fixed and what triggers a change order. If the proposal prices per use case, ask what one use case covers and what happens when it grows.
10. What happens when it breaks?
Something will break, whether it is an API change or a supplier who switches invoice layouts. Ask who gets the alert and how fast they respond, and get it in writing. Then ask whether your own team could fix it without them.
11. Which models do you use, and why?
A good answer names a model for each task and explains the trade-off between cost and accuracy. It also says how hard a swap would be. If a model vendor owns the partner or pays it, ask what they would recommend if their own model were the wrong fit.
12. Can we talk to a client who has been live for a while?
Ask for a reference whose system has been in production long enough to break and get fixed. A short call with that person tells you more than any case study.
What AI partner badges prove, and what they don't
A badge proves a firm joined a program and met its bar. Whether the firm has put a workflow like yours into production is a separate question, so read the bar before you trust the badge.
The bars vary a lot. The Claude Partner Network (March 2026) is free to join at the entry level. Its Select tier asks for 10 certified staff and 2 production customers in 12 months, and it also wants 1 public customer story. OpenAI's Partner Network (June 2026) is a $150M program with tiers from Select up to Elite, and it has not published the thresholds. Salesforce's FDE Partner Network (April 2026) admits small regional firms next to Accenture and Deloitte, and pays incentives when agents reach production.
Read those terms as a buyer. A free entry-level badge says nothing about delivery, and a Select-tier Claude partner has had at least 2 production customers in the past 12 months, a real bar but a low one. A partner that a platform pays when agents go live has a reason to put an agent in every step, so ask who else pays your partner. None of it replaces question 2.
Where a small studio like DK fits, and where it doesn't
DK Studio fits mid-market teams that have a live process to fix and a person who owns it. We sell implementation, so hold what follows to the same 12 questions.
Gianluca Boccadifuoco (CEO) and Christian Vismara (CTO) founded the studio in 2025. Christian is in New York and can be on site; the rest of the studio works remotely. Project work starts with a 2-3 week Deployment Diagnostic at $7,500 fixed, credited against the Production Sprint if you go ahead. The Sprint takes one workflow live in 6-8 weeks for $25,000-$60,000 fixed, with evals, an audit trail, human approval where money or risk is involved, shadow mode before go-live and a written handover. If you want ongoing capacity, an Embedded Engineer works inside your team 2-3 days a week for $16,000-$20,000 per month, and after launch the AI Team retainer keeps the system running from $3,000 per month. The details are on our AI implementation services and forward-deployed engineering pages.
Pick someone else when:
- You need a large team on site in several countries at once. A big integrator is built for that.
- Blackstone or Hellman & Friedman own your company. Ode was set up with their portfolio companies in mind, so talk to them first.
- All you need is a model call inside an app you already run. That is a smaller AI integration job.
- You haven't picked a workflow yet. Our five-day AI Readiness Audit, from $800, is for owners deciding what to build first.