TL;DR
- A London company that builds AI for banks is hiring a Forward Deployed Engineer and a Forward Deployed Operator at the same time. Different job descriptions, different experience bars, different interview loops. One writes code. The other finds out how the bank actually works. They are not the same job, and the company clearly knows it.
- In the past year at least six different groups landed on that second title independently, from an Instacart operations director to an industry analyst body to a Zapier executive. Every one of them defines it differently. That disagreement is the whole story.
- Your next action. Before you buy or build anything else with AI in it, write down one process the way it actually runs today, including the workaround nobody documented. If two people on your team describe it differently, you have found the real project.
A founder called us this spring about an AI tool their team had quietly stopped using. It had been live for four months and it worked exactly as sold. The problem was that it automated a three-step approval nobody had followed since 2023, so using it properly meant doing the work twice. Nobody had lied to them. Nobody had built the thing wrong.
We have been embedded inside operations running on spreadsheets and operations running on nine-figure systems, and the failure looks the same in both. The technology is almost never what breaks. What breaks is the distance between how the work is described and how the work is done, and for most of the last decade there was no job title for the person who closes it. And that is changing fast enough.
Two jobs, two interviews, one company
Tuza is a London company that builds automation for banks. Right now their careers page has two openings sitting next to each other.
The first is a Forward Deployed Engineer: “half engineer, half systems thinker.” They want three or more years as a software engineer shipping to production, TypeScript, Python, Postgres. The interview includes a two-hour programming exercise where you debug and extend one of their applications.
The second is a Forward Deployed Operator: “part product, part ops, part consultant.” They want three or more years in product operations, consulting, or banking operations. The interview is a take-home product exercise and a workshop.
Two roles. Two experience bars. Two entirely separate interview loops. No company builds a coding test for one job and a product workshop for another unless it is convinced they are different jobs.
Then there is the detail we keep coming back to. Inside the engineer’s job description, Tuza writes that you will “work closely with product, FDOs and customers.” They use the abbreviation in passing, in someone else’s job posting, the way you would write “legal” or “finance.” This is not a company introducing a new idea. This is a company that already has these people.
And read what they ask the operator to actually do: “map how onboarding, risk, pricing and servicing actually work today.” Then, under what they are looking for: “bias to action, you like shipping and iterating, not writing endless slide decks.”
We have been making that argument for two years. It is bracing to find it written down by a bank software company in Shoreditch that has never heard of us.
Six groups named the same job and described six different jobs
Tuza is not alone, and this is where it gets genuinely interesting.
In March, Patrick Balbierz, a Director of Customer Experience at Instacart who came up through critical response and trust and safety, published a piece calling it the Rear Deployed Operator. His framing is the best one we have read. He describes two professionals walking toward each other on the same street: the engineer traveling from headquarters into the operation, and the operator traveling from the operation into technical capability. AI is what let him make the second trip. In his words, neither role is complete on its own.
In April, TSIA, the research body that mid-market technology services companies pay to tell them where the industry is going, published the operator as the second half of a “service pod.” Their version: the engineer builds the first thing, the operator owns whether it keeps working. Their research lead has since built an entire operating model around the idea that somebody has to own customer value after go-live, because right now nobody does.
Last July, SuperDial, a company building voice agents for healthcare, said the quiet part out loud. Their forward deployed operators exist because the context their agents need lives “in the minds of the operators our agents work alongside.”
And last week, Philip Lakin, who built the NoCodeOps community before Zapier acquired it and now runs AI transformation there, argued that these people already existed. He calls them the spiritual evolution of the no-code operator: the figure-it-out people who were building shadow systems out of Airtable and spite long before anyone offered them a title.
An Instacart operations director. An analyst body. A healthcare AI company. A Zapier executive. A London fintech. Us.
Six groups, one title, six different definitions. That is not a coincidence and it is not a trend piece. That is a market discovering a job it has needed for years and not yet agreeing on what the job is.
Where the value actually sits
BCG’s AI Radar surveyed nearly 2,400 executives this January, 640 of them chief executives. Companies expect to double what they spend on AI this year, from 0.8 percent of revenue to about 1.7 percent. Half of the chief executives said their own job is on the line if it does not pay off.
So the money is committed. The open question is where it lands.
BCG’s own answer is a rule they call 10-20-70. Ten percent of the value comes from the algorithms. Twenty percent from the technology and the data. Seventy percent from people and processes.
The January survey found something quieter that we have not seen anyone quote. Confidence in AI’s payoff falls the further you get from the corner office. Sixty-two percent among chief executives. Forty-eight percent among executives outside the C-suite. BCG’s own reading is that the people closest to the day-to-day work may simply have a more realistic view of what it will take.
We would put it more plainly. The people who know how the work actually runs are the ones least convinced it is going to work. That is worth sitting with before you approve the next budget.
Now compare all of that to where the hiring is going. A census of live job postings dated July 4 found more than 1,200 open Forward Deployed Engineer roles across 669 companies, at a median posted base of 185 thousand dollars. Anthropic and OpenAI both launched multi-billion-dollar ventures this spring built specifically on putting engineers inside client organizations.
Search that same census for the operator side and you will not find a category. It does not exist yet.
The market has correctly identified that deployment is the bottleneck. It is staffing the 10 percent.
What we mean by forward deployed operations
So here is our definition, which is the one we would defend against the other five.
A forward deployed operator goes in before the tool. They map how the work actually happens, including the workaround that became the process and the approval step that exists because of a mistake someone made three years ago. They fix the process first. Then they build the infrastructure that fits the operation as it truly runs. Then they hand it over and leave.
We call the first part Process Archaeology, and it is the part everyone wants to skip. You cannot automate what you cannot describe. An agent will not run a process nobody has written down. It will run whatever it was told, at speed, past every exception the person who does the job in real life handles without thinking.
Our operators come from the floor, not from the codebase. That is the whole bet. Balbierz is right that the two roles are walking toward each other, and we think the trip is shorter from the operations side, because the hard-won knowledge is knowing which problems are worth solving and how the work fails in practice. AI closed most of the distance on the building.
The fair objection, and our answer
There is a real critique of this whole category and it deserves a straight answer.
Antoine Moyroud at Lightspeed put it bluntly earlier this year: used poorly, forward deployment is just expensive services with a nicer title. Gergely Orosz has watched the engineer version drift until it is hard to tell apart from a solutions architect or a consultant. Both are correct. A role defined by embedding inside a client, with no defined end, is a staffing contract wearing a better word.
Our answer is that the exit is designed on day one. We run three phases. A Process Archeology, where we map what is really happening against what everyone believes is happening. Deployment, where we fix the process and build the systems around it. Then the Leadership Tail, where we hand the controls to your team and step back until they run it without us.
Success is not a renewal. Success is that you stop needing us. Any definition of this role that does not include leaving is describing a different job.
Three questions to tell whether you need this
Not a hiring checklist. A diagnostic on whether your operation is ready to be automated at all.
1. Can two people on your team describe the same process the same way? Ask them separately. Write both answers down. If they differ, no tool will settle it, because the disagreement is the project.
2. When the last tool went in, did the old workaround disappear, or did it move? If your team still keeps the real numbers in a side spreadsheet, the system was installed on top of the process instead of into it.
3. Who owns the outcome ninety days after go-live? If the honest answer is the vendor, or nobody, you have bought a deployment and not a result.
Three yeses and you are ready for tooling. Any no, and the tool will land on the same ground the last one did.
The role has been real for years. What changed this year is that enough people needed it at once to start considering what to call it.
Next Tuesday, that argument stops being an essay. Philip Lakin, one of the six voices above, joins me for the first episode of The Forward Deployed Operator, a weekly show about the operational reality behind all of this. Thirty minutes on when to automate and when to go rebuild the process first.Subscribe to the podcast
OPS INTEL
↑ Dry van spot rates top contract rates for the first time since February 2022 (DAT Freight and Analytics, July 9). Van spot hit 3.00 dollars a mile in June against a 2.89 contract rate, and flatbed set an all-time high at 3.69. Rates rose faster than volume, so this is capacity tightening rather than a demand surge. When spot runs above contract, your carrier’s contract rate is the cheap option, and every lane you still tender to the spot market is where the freight budget breaks first. Reefer at 3.39 hits hardest if you ship chilled.
👀 July imports forecast to set an all-time record, then fall off a cliff (National Retail Federation and Hackett Associates, July 8). The 2.47 million TEU forecast for July would beat the record set in May 2022. Then August turns negative and September lands 5.7 percent below last year. Everyone pulled fourth-quarter inventory forward at once, so port and drayage congestion peaks right now, and whoever guessed wrong on the frontload will be discounting into your category all fall.
↓ Manufacturing input prices post their steepest one-month drop since 2022 (Institute for Supply Management, July 1). The prices index fell 9.1 points to 73.0 and supplier deliveries loosened by 3.2 points, though this is still the twenty-first straight month of rising input costs. Treat it as a window, not a turn. This is the moment to reopen the ingredient and packaging quotes you signed at the May peak.
↑ Canadian manufacturing sales hit a record, and so did unfilled orders (Statistics Canada, July 15). Sales reached 78.1 billion dollars in May, up 13.4 percent year over year, while unfilled orders set their own record at 131.5 billion against an inventory-to-sales ratio of just 1.61. Quebec was the one province that moved backwards, down 0.8 percent. Your contract manufacturer’s queue is now longer than their stock can cover, which makes the quoted lead time for a fall production run the number to renegotiate this week.
👀 Two-thirds of Canadian firms are absorbing higher costs instead of repricing (Bank of Canada Business Outlook Survey, July 6). Nearly three-quarters reported costs rising from the Middle East conflict. Of those, roughly 40 percent are absorbing it entirely and another quarter are passing on only part of it. The share of firms planning around a recession nearly doubled, from 9 to 17 percent. Margin leaves quietly in exactly this pattern, so run the pass-through math on your own freight and energy surcharges before the next quote goes out.
👀 HelloFresh went from 100 to 500 chilled SKUs in the same square footage (Supply Chain Dive, June 30). A 13-robot pilot grew to 39 robots inside three months, and the change that made it work was hardware: a heated motor so the robots run continuously in cold storage. Worth reading next to this issue, because the win was not headcount. It was running five times the SKU range through a cold room they were already paying rent on.
Ysi
Vantelira Inc.