Forward deployed engineers or AI consultants?
Not long ago, this was easy to explain. A forward deployed engineer wrote production code inside your company. An AI consultant told you what to change and left a deck behind.
That line has mostly gone. Accenture and Deloitte now sell forward deployed engineers. McKinsey says the term “may be a buzzword” for something firms have offered for years. OpenAI’s own announcement calls Tomoro an “applied AI consulting and engineering firm” and its people forward deployed engineers in the same paragraph.
The basic answer still holds. If you can name the workflow and need a working system, hire engineers. If you cannot name it yet, or the barrier is how your company is organized, hire advisers. If the system will be core to your business for a decade, build your own team.
What has changed is how you tell them apart. Titles no longer help. Who the people work for, what they are paid to sell next, and what you own when they leave do. Lunon sells engineering, so read this with that in mind. We have kept the evidence that cuts against us.
A term with a stricter meaning than most people use
Palantir made the role famous, internally calling these engineers Deltas. Its 2019 explanation is still the clearest: a developer focuses on “one capability, many customers,” a Delta on “one customer, many capabilities.” The same post draws the line against consultants. “The critical difference is that we are actually deploying existing software products.”
By that test, most firms using the title today would fail, ours included. We have no product to deploy. We use the way of working. The market has not waited for the definition: Lightcast counted about 200 US postings in 2024 and over 5,200 in the first seven months of 2026.
Three kinds of engineer, three sets of incentives
Vendor engineers work for whoever sells the product or model: Palantir, AWS, OpenAI. What they learn goes back to the vendor, by design. “Our goal is for each engagement to leave the next customer with a better starting point,” OpenAI writes.
Consultancy engineers come from a services firm, usually working on a partner’s platform. Accenture has programs with Microsoft, Google Cloud, ServiceNow and Workday. They have a reason to sell the next phase.
Independent engineers come from a firm with no product to sell. That removes the pull toward one platform and adds a different one: a reason to extend the engagement. Ask any of the three what happens when the work is done.
What the rate cards show
Almost nobody comparing these models cites a real price. Government price lists are public, so we pulled them. Palantir’s UK list shows implementation and engineering at £175,000 per person per quarter, about £2,700 a working day by our arithmetic, before licenses or usage. The table below comes from the US government’s rate database. These are ceilings for federal buyers, with categories each vendor defines for itself.
US federal ceiling rates per hour for selected engineering and advisory roles, as of 6 November 2026. Ceilings, not prices paid, and categories are defined by each vendor.
Source: US General Services Administration CALC rate database, pulled 6 November 2026.
Two things stand out. The usual assumption that advice costs more than building holds only some of the time. BCG’s ceiling for a strategy associate with two years’ experience sits above its principal data scientist with seven. But on Deloitte’s January 2026 UK list, it runs the other way: a lead developer is £2,020 a day and a lead business analyst £1,840. And neither model is reliably cheaper. Palantir’s UK rate, about £2,700 a day, is above every rate we found on Deloitte’s list, while smaller firms advertise $7,000 to $40,000 per engineer per month, by their own claims.
The billing myth
A common line says consultants bill hours while forward deployed engineers get paid on outcomes. The public record does not support it. McKinsey says about a quarter of its fees are performance-based. Accenture’s annual report says many contracts tie its pay to outcomes. Palantir’s published UK price is per person per quarter. Ask for the terms in writing.
What the evidence can and cannot tell you
We looked for a study comparing the outcomes of forward deployed engineering and consulting. We did not find one. Anyone who says one model is proven better is guessing, and that includes us.
The famous numbers are weaker than they look. The MIT NANDA report behind “95 percent of AI pilots fail” rests on 52 interviews and 153 survey responses, calls itself preliminary, and its own chart implies about a quarter of pilots reached implementation. RAND’s report attributes “more than 80 percent” to “some estimates.”
What seems to predict success has little to do with who delivers. RAND’s interviews put misunderstanding the problem first among causes of failure. In McKinsey’s 2026 survey, about three quarters of high performers fundamentally redesign workflows, against about a quarter of others. The hard part stays hard whoever you hire.
The case against engineers
Lock-in is the obvious one. Gartner warns that relying on vendor engineers for every change “creates enormous lock-in,” and advises clients to “figure out the exit strategy.” Whatever the engineers learned about your business leaves with them unless you plan for it.
The economics are the stronger one. a16z argues that bespoke deployment does not pay for an 8 percent gain in a mid-market sales workflow, and it is right that the workflow has to matter. Palantir could absorb early losses because software revenue followed: its 2020 filing shows a $65.4 million contribution loss in its earliest customer phase. A services firm has to charge for the work.
Custom code also ages. OpenAI reports removing roughly 90 percent of the custom code from each project as its models improved. Ask what you are paying for, and whether it will last.
Which do you need
This part is our judgment, because the evidence does not settle it.
Hire advisers when the question is what to do, when the barrier is your operating model, or when the decision needs board-level cover. Even OpenAI brings in McKinsey and BCG for strategy and change management around its own engineers.
Hire engineers when you can name the workflow, the data sits across several systems, and nobody on staff can wire it together in production.
Build in-house when the system is a permanent core competency and you can hire for it. Levels.fyi data puts median US total compensation for the role at $216,250, before overhead.
Buy software when a product already does the job. Plenty of workflows do not justify custom engineering.
Seven questions to ask before anyone starts
Who employs them, and what are they paid more for selling next?
Do they get write access to production, or read-only exports?
On the day they leave, what do we own: code, evaluations, runbooks, prompts, know-how?
What may they reuse at our competitors?
Who on our side takes over, and when do they start shadowing?
Which model or platform does this commit us to, and can we change it?
What will it cost to run after they go, including tokens?
Get the third one in writing. In the US, code written by a contractor generally needs a written assignment for you to own it, on our reading of Copyright Office Circular 30. Confirm it with counsel.
Where Lunon stands
Two of our engineers work inside your company from the first workshop through full rollout, build on the systems you already run, and route each step to whichever frontier model is strongest at it. You own the system, the data and the know-how. If you are still comparing firms, our guide to AI transformation companies covers sixteen, including who owns them and where each falls short. If you can already name the workflow, let’s talk.
Rates are public-sector figures from Deloitte’s and Palantir’s G-Cloud 15 lists and the GSA database in November 2026, and are not directly comparable. Sources checked 6 November 2026. Nothing here is legal advice.
