Forward-deployed, for companies Palantir will never call

Forward-deployed engineering is an old idea with a new budget behind it. Instead of selling software and leaving the customer to make it work, the vendor sends an engineer into the customer's operation. The engineer sits with the people who do the work, learns where the data is wrong and where the rules live in someone's head, and builds until the software runs there. Palantir made the model famous by doing exactly that inside governments and large companies.

The model works because most failed software projects fail on the last mile. The tool is fine; the data, the exceptions and the approval rules of one specific company are not. An engineer in the room closes that gap in weeks. A support queue takes months, or never does.

The largest names in enterprise technology adopted it during 2026. Amazon committed $1B of internal resources to a new forward-deployed engineering organisation on 30 June. Databricks launched a forward-deployed engineering organisation on 11 June. Accenture launched a forward-deployed engineering practice with Microsoft in March. TechCrunch reported a Christian & Timbers study projecting demand for forward-deployed engineers to rise 2,100 percent by the end of the year.

None of this is a coincidence. AI agents are the clearest case yet of software that only works once someone fits it to a real workflow. A model can read an invoice anywhere. Knowing which invoices your controller wants flagged, and why, takes a person who has watched your close.

Two things changed at once. Models became good enough to read messy documents and data without a rule for every case. And that made the last mile the whole job, because the remaining work is exactly the company-specific part no product team can build in advance.

The catch is the price. Forward-deployed engineers are among the most expensive hires in technology. A program built on those salaries, plus the vendor’s margin, fits a bank, a hyperscaler’s largest customers or a government. It does not fit a company with $10M to $200M in revenue, a controller and a small finance team.

Those companies have the same problem the large ones do. Their close takes too long, their matching is manual, and their reporting arrives late. They have the data and the approval rules. What they lack is someone to fit the agent to the workflow and stay while it settles. The market has built the model for the top and left the middle with self-serve tools and a help centre.

Our Stay stage keeps the parts of forward deployment that matter and drops the parts that make it expensive. The engineer who built the agent runs it. We sit in the close meeting during the pilot. We watch the exceptions ourselves, fix the rules, and rerun the evals every month. When the next workflow is ready, the same engineer scopes it.

What we drop is the permanent on-site team. A mid-market workflow does not need five engineers for a year. It needs one engineer who knows the system, a two-week audit to learn the workflow, a six to eight week pilot to build it, and a monthly retainer to keep it right. Priced that way, the pilot starts at $20,000 and the retainer at $5,000 a month.

We also drop the idea that the vendor owns the result forever. The contract carries an exit plan: the runbook, the code, the credentials and deletion of our copies. Forward deployment at the top of the market often becomes a dependency. For a smaller company, the goal is a system its own team can run when it is ready.

The engineer also carries the context that usually gets lost. When most software rollouts close, the people who learned the workflow leave, and the knowledge goes with them. Keeping the same engineer on the retainer means the reasons behind each rule stay in the room, and a change in the business becomes a change in the agent within days.

If you are buying agents without a forward-deployed budget, look for three signs that a vendor works this way. First, an audit before any build, so the agent is shaped by your workflow rather than a demo. Second, a named engineer who stays after launch, not a handoff to support. Third, an eval you can read every month, so you know the agent is still doing the job.

A vendor that offers all three is doing forward-deployed work at a price the middle of the market can pay. A vendor that offers a login and a help centre is selling a tool, whatever the pitch calls it.

Price is the last check. A vendor doing this work properly will quote a fixed price after the audit, not a day rate. Day rates are how forward deployment is sold at the top of the market, and they are how it becomes unaffordable in the middle.

Borrowing the method is not the same as matching the scale. One engineer cannot run fifty workflows at once, and we do not pretend to. We queue workflows one at a time, and we tell a client when their volume needs more than a retainer covers.

We are also clear about what does not fit. A workflow with no stable approval step, or data nobody can give us read access to, is not ready for an agent, and no amount of forward-deployed effort changes that. The audit exists partly to say so early, before a client has paid for a build.

The workable version of forward deployment for the mid-market is narrower than the Palantir version. It is also the only version most companies can buy, and for the workflows that eat a finance team’s week, it is enough.