Three versions of the same project
Two problems surfaced in a recent demo of our company brain for client delivery teams.
A grocery customer's scope still describes an older version of the product. Engineering has since removed a capability that the scope relies on. The document sits in one system; the product decision sits in another. Nobody has joined them up.
At a pet retailer, sales has promised touchless invoice posting. The platform requires a customer approver to release each posting. The promise is in a customer conversation. The limit is in the delivery and platform records. The same missing context can delay customer onboarding.
Both companies are fictional. The demo uses Fieldwork AI, an invented AI company that sells through forward-deployed engineers. In the demonstration I ran, the company brain flagged both conflicts and cited the evidence on each side. A connector-based assistant in the comparison repeated an outdated scope without finding the conflict.
That is one demo. We haven't run a controlled benchmark of assistants. It illustrates the gap I keep seeing in real delivery teams: what was sold, what was scoped and what was shipped drift into different accounts of the same project.
Hiring FDEs is easier to count than the learning they bring back
Forward-deployed engineers work alongside customers to make a product useful in their environment. The role is growing fast. In his November 2026 job-market report, Lenny Rachitsky puts open FDE roles at about 1,500, a 4.6-fold increase over the preceding year. He notes that some of the growth may be sales-engineering roles being renamed. Openings measure advertised demand. They don't count hires.
In a conversation shared by David Senra, ElevenLabs' Mati Staniszewski describes FDEs as part of product. Solving the customer's problem is one job. Bringing what they learn back into the product is another, and teams often miss that second job.
The practical test is whether each deployment makes the next one easier. That depends on four handoffs:
- Sales to FDE: what the customer was promised, against the time, budget and capability available.
- FDE to product: what delivery learned and which work should become reusable product.
- Product to sales: what shipped, changed or was dropped before the next proposal is written.
- Product to FDE: those same changes reaching engineers who are already mid-delivery.
A company can employ excellent people at every point and still lose context between them.
The usual fixes are good. They need keeping current.
The published playbooks put ownership in the right places. Falk Gottlob's rule is that sales owns the number and the FDE owns scope. Before signature, promises are sorted into what ships today, what needs building and what cannot be committed. Commercial and technical owners agree on the same page.
PostHog's handbook asks for a measurable outcome, rough timeline and shared definition of done, confirmed with the customer before work starts. A deliverable that doubles in scope needs a new quote. Its product-engineering guidance tells FDEs to file issues with the owning team and make the customer context explicit.
These are useful controls. The difficulty is keeping the facts behind them current. A scoping page can be right when signed and wrong a week later. A product readout carries what somebody remembers to raise. The rest stays in transcripts, threads and coding sessions.
The process still needs owners, agreed scope and regular review. It also needs a way for a later decision to reach every promise and project it changes.
What the drift looks like in client work
The following observations come from two engagements we are working on. Both clients are anonymised; these are field observations rather than a representative study.
The hard thing gets sold while the easy thing sits on the shelf
The first is a Series A AI company of about 60 people. Each account involves scoping and forward-deployed engineering. Sales sometimes commits to work the FDEs cannot deliver inside the agreed time or budget. Meanwhile, capabilities built for previous customers go unsold because the commercial team cannot see them.
Engineering receives a difficult deadline while knowing an easier route existed, one the customer might have accepted. The incoming COO described the cost: today's decisions can invalidate yesterday's understanding, and even one person spending a day on the old understanding hurts.
Two pods finish their work and discover a gap between them
The second client is a 20-person AI startup working across two cities. One pod built invoice capture and validation. Another built purchase-order matching. Only after both finished did they discover a step between validation and matching that nobody owned.
An engineer told us an agent reading the design discussions and transcripts could have surfaced the gap earlier. Another wanted relevant discovery-call context as soon as the commercial team left the customer, before an impractical commitment hardened into a plan.
The work is spread across sessions as well as teams
The operations lead described engineers running 10 to 15 sessions at once. Each session has its own context; a coding branch can stay isolated until merge. Sales, delivery and engineering also use different assistants. The decisions made in those sessions often never reach the next person working on the same customer.
This is the alignment problem inside a delivery organisation. Each person moves faster, while the team spends more time discovering what the others already learned.
Two fair objections
Fix the product, rather than relying on permanent FDE effort
I agree with the direction of this objection. Rocio Wu at F-Prime warns about custom delivery becoming the default. Her threshold of 30% to 40% of deployments needing significant FDE effort is a rule of thumb from practice. It won't fit every company.
Marc Andrusko at Andreessen Horowitz makes a related case for time-boxing early deployments and converting bespoke work into reusable product on purpose. A workaround that stays in a private session cannot help the next deployment. Capturing learning supports the product transition; it does not remove the need to make it.
Keep the status, instead of reconstructing it every morning
In a conversation reported by Christophe Barre, Russell Garner of Omni challenges the habit of sending an agent through every source just to recover the latest status. A good process should keep that status from being lost.
That is the right goal. A company brain keeps an account of the work up to date as connected sources change, so anyone asking later gets the current status without rebuilding it. It is not instant: a message written a minute ago may not be reflected yet. The stand-up pilot below tests how well that account holds up.
Connectors provide access. The joins still matter.
Connecting an assistant to each tool is useful. It does not, by itself, establish that a scope document, a sales call and a product decision concern the same commitment.
Some tools support semantic search and some assistants discover tools on demand, so it is too broad to say connectors only match words or always load every tool description. The practical questions are whether the system can tell that a scope document, a call and a ticket concern the same customer, spot evidence that conflicts, and say which decision is current.
Separate searches can still miss a commitment that a transcript phrases in other words. An early result can be mistaken for a complete answer. Large tool catalogues and raw source documents can consume context before the agent reaches the decision. We discuss that cost in The $10,000 analyst.
Decision files in a repository help too. They become much more useful when they have owners, dates, explicit replacement links and access rules, and when they connect to the customer conversation that explains them.
What this lets a delivery team do
The company brain gives authorised people and agents one connected view across sales, delivery, product and engineering. Anyone working on an account can see the commitments, risks, blockers, scope changes and decisions that affect its projects, whichever team's system they were recorded in.
See when a promise and a product decision conflict
When a later product decision undercuts a sales commitment, the mismatch is flagged for review with the sources on each side. The scoping page stays useful as the product changes. The brain can only catch what it can see: a conversation that was never connected remains a gap.
Keep approved design decisions after the session ends
Tian Pan describes agents reopening questions whose original reasoning was never preserved. Duncan Davidson describes the opposite failure: an agent enforcing an obsolete decision because the record still presented it as mandatory.
When a change merges, the human design decisions behind it are kept: who approved them, what changed, why, when and which evidence supports them. A later decision marks an earlier one as replaced, so the next person or agent to pick up the work starts from the decisions still in force, within their own access rights.
A line in the shared AGENTS.md can set that expectation, but the capture itself has
to be built into the merge and checked.
We work this way ourselves and have found it makes collaboration easier. With the second client, decisions are being captured when a change merges into the preview environment. A merge shows that code changed. Reaching production is a separate event, with its own evidence.
Keep the reasons behind the ticket
A ticket compresses a conversation. It can lose the hesitation, workaround or customer constraint that made the request important. With the ticket linked to its source, the next engineer or agent can go straight back to that reasoning. Our own tickets live in the brain, alongside that evidence.
Preserve permissions across every handoff
Everyone sees only what their access to the original source allows. Shared channel history stays limited to people with channel access. Direct messages and private personal context stay outside company search. The permissions boundary matters just as much when the asker is an agent.
A daily stand-up as an alignment test
The second client chose a concrete test. Each morning, a person gets a draft stand-up: yesterday's work, today's plan, blockers and dependencies. They edit and approve it. Only the approved version is written back.
The pilot lead wrote nine evaluation checks before the trial. The agreed acceptance target was at least eight passing on three consecutive days. We will report the result when the pilot finishes.
Editing effort gives us one signal about whether the brain has useful context. Few edits do not prove the draft is correct: people can overlook mistakes. The evaluation checks and source review still have to establish whether its claims are supported.
What we have learned so far:
- The stand-up needs several sources. Tickets, meetings, chat and calendars each hold part of the work. Agent sessions can add context where access and data-handling arrangements permit.
- Tickets alone are insufficient. Some engineers seldom update them, and much of the reasoning sits in coding sessions.
- Noise can matter more than volume. In this pilot, an automated alerts channel with about 1,600 messages made answers worse until we excluded it.
- Latency affects adoption. We cut observed query times by about half during the trial. That figure comes from this trial alone.
The team asked to extend the trial, build a messaging bot around the stand-up and move the brain into its own cloud. Those are encouraging adoption signals; they are not the acceptance score.
The hard parts remain human as well as technical
Sessions can contain customer data. We agree source access, contractual terms, which model providers may handle the data, and hosting requirements before any raw sessions are connected. A deployment inside the client's own cloud can support its data boundaries. Until those arrangements are ready, approved decision records can capture useful learning without importing full session transcripts.
Company identities need care. A nickname may be clear in one person's context and ambiguous across a whole organisation. The early pilot surfaced duplicate records for the same company that we are still cleaning up. A weak match should prompt review rather than a confident merge.
Capture changes behaviour. People may move sensitive conversations elsewhere if they expect everything to become public. Access should follow legitimate roles and source permissions. A useful company brain does not require everyone to see everything.
The brain cannot set direction. If leaders have not decided what matters, it cannot infer the right priorities from a pile of activity. It helps people stay aligned to decisions people have made.
Start with one handoff you can test
- Map the four handoffs. Record where promises, scope, product changes and delivery learning live, and who owns each transfer.
- Choose a concrete measure. Track commitments corrected before delivery, time spent reconciling scope, repeated implementation work or the accuracy of a reviewed stand-up. A completed deployment should also leave reusable work and product insight.
- Capture decisions at merge. Make sure every merged change leaves behind its approved reasoning and sources, and check what gets kept. Keep the decisions rather than every raw session.
- Expand after the first loop works. Sales and the FDE team are a useful starting pair. Bring in further departments as source access, ownership and checks are ready.
We install a first company brain use case in two weeks, with scope and access agreed before work starts. Choose 60x Cloud with EU or US data residency, or Your Cloud inside your own Azure, AWS or GCP cloud environment. Merge-time decision capture and further workflows are scoped with the use case.
Book a demo and bring a project where the promise, scope and delivered product have drifted apart.
FDEs learn things in the customer's environment that nobody else can. The value compounds when that learning reaches sales before the next proposal and product before the next release.
Sources
- Lenny Rachitsky, State of the tech job market: late 2026, 10 November 2026. Openings data and the role-renaming caveat.
- David Senra's conversation with Mati Staniszewski, shared September 2026. The product-learning role of forward deployment.
- Falk Gottlob, In the Room, Not on the Quota, 29 September 2026. Commercial ownership, scope and the pre-contract agreement.
- PostHog, How forward deployed engineers work, and Working with product engineering, checked 15 November 2026.
- Marc Andrusko, The Palantirization of everything, Andreessen Horowitz, 16 January 2026.
- Rocio Wu, The Uncomfortable Truth About FDEs, F-Prime, 6 February 2026. The deployment-effort threshold is a rule of thumb.
- Christophe Barre's account of Russell Garner's status-update critique.
- Tian Pan, The ADR Your Agent Never Read, 2 July 2026.
- Duncan Davidson, Coding Agents Love Decision Records, O'Reilly, 2 October 2026.
Client observations are the author's first-hand account and are anonymised. Fieldwork AI, Grocer A and Pet Retailer B are fictional. Illustrations paraphrase scenarios and workflows; they do not show real client interfaces or data.