AI tools for Forward Deployed Engineering

The Next Bottleneck Is Understanding the Business

A customer spent $5 million and five years migrating to NetSuite and has zero appetite to rip anything out. That single constraint shapes an entire architecture — and the talk's most useful admission is that frontier models are bad at the one thing this work depends on.

Video thumbnail
📺 AI Engineer ⏱️ 20:22 📅 28 July 2026
Forward deployed engineering Enterprise AI Systems of record Knowledge graphs Post-training Process automation

⚡ Execution stopped being the constraint 1:07

The talk opens with a before-and-after that most people in the room could verify from their own week. Two years ago, "has AI completed an end-to-end task for you?" would have drawn almost no hands. Today it draws all of them.

So the argument is that intelligence and tooling have stopped being the limiting factor. What remains is something less tractable:

"The bottleneck that is still here is how much can you understand the business — because every business, every consumer is different."

The example given is precise: a sales department at a healthcare company operates nothing like a sales department at a SaaS company. The knowledge that distinguishes them lives in people's heads, and it does not survive being pasted into a prompt.

The framing worth carrying away, independent of the vendor: when capability becomes abundant, the scarce input becomes specification. Knowing what should happen — including the exceptions nobody wrote down — is the part no model can supply for you, because it is not in the training data. It is in Sarah's head in accounts payable.

🗺️ What the documentation never says 3:27

The first job of a forward deployed engineer, as described here, is mapping how the work is actually done. Engineers embed with a single department — finance, say — and interview the process leads for AP, AR, card reconciliation, banking, billing and FP&A.

And the emphasis is not on the happy path:

"A lot of the documentation you have at companies is about the golden path and maybe an edge case or two. But when we talk to customers, it's a lot of: Sarah in AP handles the workflow today in this way, but when things go wrong she sends it over to Chris, who then takes four days of cycle time to handle reconciliations between a purchase order and an invoice."
That example is doing a lot of work, and it is worth unpacking. The four-day reconciliation is not in any process document. It is not in the ERP. It exists as an informal handoff between two people, and it is simultaneously (a) where the cycle time actually goes and (b) invisible to anyone evaluating the process from a system export. Any automation built from documentation alone will model a process that does not exist.

Anyone who has tried to automate a real back-office process will recognise this immediately. The documented workflow is the workflow on a good day; the exceptions are where the labour is.

💀 Why pilots die 4:54

The second job is re-engineering the process around AI rather than bolting AI onto it. The diagnosis offered for the industry's poor hit rate is that "AI is being slapped on top of broken processes".

The most interesting part is the prescription, because it argues against maximum automation:

Of an eight-step processHandling
Four stepsFully autonomous
Three stepsHuman in the loop
One stepHuman only — risk too high, or too little value in automating

And the reasoning behind the restraint is about adoption, not capability. Replace an eleven-step workflow with a one-step one and the users may reject it outright — but change too little and there is no return to capture. The design target is the narrow band between those failures.

This is the most practically useful idea in the talk. Automation projects are usually scoped by what is technically automatable. This scopes by what the organisation will actually adopt — and treats the residual human steps as a deliberate design choice rather than an admission of incomplete work.

🏛️ The $5 million constraint 7:16

The third job is deployment, and here the constraint is stated as a direct customer quote: they spent $5 million and five years migrating to NetSuite.

"If you're telling them, hey, I have this fancy AI tooling, but by the way you have to migrate off of NetSuite — they're going to tell you to get out. They don't have any appetite for that."

So the architecture is agents deployed on top of existing systems of record — Salesforce, NetSuite, Dynamics, SAP — with governance and evals in the vendor's own platform layer, and no migration asked of the customer.

Whatever you make of the vendor, the constraint is real and under-served. A great deal of enterprise AI tooling implicitly assumes the data can be moved somewhere more convenient. For organisations whose systems of record represent a decade of integration, compliance sign-off and audit history, that assumption disqualifies the tool before the evaluation starts. "Meet the systems where they are" is a genuine architectural position, not just a sales line.

🧑‍💼 The hiring problem 8:28

Everyone declares 2026 the year of the forward deployed engineer. The talk's contribution is naming why that is harder than it sounds: the role needs someone in the technical top percentile and capable of sitting with a finance process lead and extracting undocumented knowledge in real time.

The usual paths — train a consultant on the technology, or train an engineer on the soft skills — both produce someone strong on one axis. The stated conclusion is that this combination does not scale by hiring, which is what motivates building an agent to amplify the engineers you already have.

The description of the job itself is vivid and rings true: clients emailing 24/7, hundreds of pages of documentation, and every process lead pulling in a different direction because AP depends on AR depends on reconciliation depends on FP&A, each convinced theirs is the priority.

🔧 Three stages of the FDE agent 12:18

The head of engineering takes over, and his account of the origin is refreshingly unpolished: the platform team was comfortable while the FDE team across the office was visibly exhausted. Asked about their process, the answer was uploading roughly 150 pages of documentation to Claude, waiting two minutes, and receiving analysis that was "verbose and incorrect".

StageWhat it doesStatus
1. Engagement agentPulls in meeting notes, synthesises documentation, reads slides. Answers "who owns this process?" and "is this Sarah the same Sarah?"Built
2. Workflow agentThe same agent embedded in the platform, flagging missed edge cases while the engineer builds the workflowBuilt
3. Autonomous assistantReceives client change requests by email, queries the model of the company, ships the workflow change without the engineerNot yet

Underneath everything is a single source of truth representing how the company works — implemented as a dependency graph. The observation offered about enterprise workflows is a good one: they are "remarkably linear, they just have a lot of cycles in them", and process owners want dependency ordering so that person C is not handed work before A and B have approved it.

"There were five companies downstairs trying to sell you a graph DB, and you can just use Postgres."

A small moment of candour that says more about the state of the tooling market than the rest of the talk.

🎯 Where frontier models fail 16:06

This is the most valuable section, because a vendor talk had no obligation to include it.

Asked whether Claude produces good output given properly extracted context, the answer is a flat no — described as "kind of surprising". The failure is not reasoning. It is editorial judgement:

"I only started believing in consultants once we started hiring them. They're so good at figuring out what is the part of the detail the client actually cares about and what can get glossed over. And frontier models have absolutely no concept of this."
That is a precise and unusually honest diagnosis. Long-form analysis is not a reasoning benchmark — it is a relevance-ranking problem. A model that includes everything true is exhaustive and useless; the skill is discarding true things that do not matter to this reader. Frontier models are optimised toward completeness, which is exactly the wrong bias for consulting output.

Their response is post-training on open models, with Kimi K2.6 named as a favourite — while noting that several would do fine. The goal is the balance between detail and clarity, not raw capability.

🕸️ Traversing the graph 17:16

The second half of the problem is retrieval: a large knowledge graph is only useful if the right context can be found reliably. Their approach is an RL environment with custom graph-traversal tools, including:

  • Entity resolution — confirming person A and person B are the same person. "There are a lot of Mikes in every company we work with, and Claude gets very confused by this."
  • Structural validation — identifying redundancy cycles and DAG violations in the graph
Entity resolution is an unglamorous problem that quietly breaks a lot of enterprise AI. Three Mikes, an email display name, a Slack handle and a surname in an ERP record can be one person or four. Get it wrong and the process map is wrong in a way that looks entirely plausible — the failure is silent, and it propagates into every downstream decision. Building explicit tooling for it, rather than hoping the model figures it out, is the right instinct.

🧭 Reading this as a buyer 18:37

The closing is explicitly a pitch — the company is hiring, and looking for clients. That does not invalidate the content, but it should calibrate how the numbers are read.

The ROI figures are the part to treat with most caution. Departmental transformations delivering "25%, 50%, 75%" are stated without a baseline, a measurement method, a time window or a named client. Compare that with the specificity elsewhere in the talk — $5 million, five years, four days of cycle time, 150 pages — and the contrast is informative. The operational details are concrete; the outcome claims are not.

What survives the discount is the diagnosis rather than the results:

  • Undocumented exception handling is where process time actually goes
  • Systems-of-record gravity is a real architectural constraint, not an excuse
  • Partial automation with deliberate human steps is a design choice, not a failure
  • Frontier models lack editorial judgement on long analysis — a capability gap, not a prompting problem
  • Entity resolution fails silently and needs explicit tooling

Those are testable against your own organisation without buying anything.

🔍 Claims checked

ClaimResult
MIT: 95% of generative AI pilots fail to reach productionReal — MIT NANDA, "The GenAI Divide: State of AI in Business 2025"
Kimi K2.6 as a post-training basemoonshotai/Kimi-K2.6 exists on Hugging Face
Varick Agents is a real companyvarick.ai and varickagents.com both resolve
Enterprise workflows are linear with cyclesConsistent with how dependency-driven approval chains work in practice
Departmental ROI of 25–75%No baseline, method, timeframe or client named — unverifiable
"Knowledge work is almost entirely solved"Rhetorical. Contradicted three minutes later by his own engineer saying frontier models produce verbose, incorrect analysis
An important qualification on the 95% figure, which the talk repeats without context. The MIT NANDA report is based on 52 organisational interviews, surveys of 153 senior leaders, and a review of 300+ publicly disclosed initiatives. It is a preliminary, non-peer-reviewed report, and its methodology drew substantial criticism when it circulated in August 2025. The figure is real and widely cited — but it is a directional finding from a modest sample, not a measured industry failure rate. It has become a rhetorical fixture in vendor decks precisely because it is convenient.
The internal contradiction is worth noticing rather than dismissing. The CEO says knowledge work is almost entirely solved and the remaining problem is business understanding. His head of engineering then explains that frontier models produce analysis so poor they had to post-train their own. Both cannot be true. The engineer's account is the more credible one — it is specific, unflattering, and describes work they actually did.

Verified 28 July 2026. This is a conference talk by a vendor; treat outcome claims as marketing and operational detail as testimony.

💡 Key takeaways

  1. When capability becomes abundant, specification becomes scarce. Knowing what should happen — including the exceptions nobody documented — is the input no model can supply, because it is not in the training data.
  2. The documented process is the process on a good day. The real cycle time hides in undocumented handoffs — Sarah passes it to Chris, and reconciliation takes four days. Automation built from documentation models a workflow that does not exist.
  3. Deliberate partial automation beats maximum automation. Four steps autonomous, three human-in-the-loop, one human-only. Scoped by what the organisation will adopt, not by what is technically automatable.
  4. Adoption is the real constraint on process change. Replace an eleven-step workflow with one step and users reject it; change too little and there is no return. The design target is the narrow band between.
  5. Systems-of-record gravity is an architectural fact. $5 million and five years into NetSuite means migration is off the table. Tools that assume data can move somewhere convenient disqualify themselves before evaluation.
  6. The FDE role does not scale by hiring. It needs top-percentile technical skill and the ability to extract undocumented knowledge from a finance lead in real time. Consultants trained on tech, or engineers trained on soft skills, each cover one axis.
  7. Frontier models lack editorial judgement. Not a reasoning failure — a relevance-ranking one. Knowing which true details to discard for this reader is the consulting skill, and models optimised for completeness have the wrong bias.
  8. That admission is the most credible moment in the talk, because a vendor presentation had no incentive to include it — and it directly contradicts the CEO's claim that knowledge work is solved.
  9. Entity resolution fails silently. Three Mikes, a Slack handle and an ERP surname may be one person or four. Get it wrong and the process map is plausibly wrong everywhere downstream.
  10. You can just use Postgres. A dependency graph is the useful representation because enterprise workflows are linear with cycles; the storage engine is not the interesting decision.
  11. The specificity gap is the tell. $5 million, five years, four days, 150 pages — all concrete. "25%, 50%, 75% ROI" — no baseline, no method, no client. Read the operational detail as testimony and the outcome claims as marketing.
  12. The 95% statistic deserves its own caveat. Real, from MIT NANDA — but a preliminary, non-peer-reviewed report built on 52 interviews and 153 survey responses. Directional, not a measured failure rate, and it has become a vendor-deck fixture because it is convenient.

🔗 Resources & links

🕐 Timestamp index

0:12Introduction — Varick Agents
0:55Why it is called "the next bottleneck"
1:07AI has solved the execution of work
1:52Understanding the business is what remains
2:42Operations centred on humans, then on AI
3:27What a forward deployed engineer does
4:08Sarah, Chris, and four days of cycle time
4:54Re-engineering the process around AI
5:11The 95% and 87% statistics
5:56Non-technical operators are not engineers
6:24Different enough to matter, familiar enough to adopt
6:39Four autonomous, three assisted, one human
7:16Enterprises married to their systems of record
7:41$5 million and five years migrating
7:58Building on top, not asking them to move
8:282026 as the year of the FDE
8:51Why the role is so hard to hire for
9:44Clients emailing 24/7, every lead pulling differently
10:282024 — when execution was still the bottleneck
11:09Point solutions versus department-wide
11:36The 25/50/75% ROI claim
12:18The engineering side takes over
12:44150 pages into Claude, verbose and incorrect
13:07Stage one — the engagement agent
13:27Are these two Sarahs the same person?
13:48Stage two — the workflow agent
14:35Stage three — the autonomous assistant
15:21A single source of truth
15:34You can just use Postgres
15:46Linear workflows with a lot of cycles
16:06Does Claude give good output? Honestly, no
16:38Why consultants are good at what models are not
16:53Post-training on Kimi K2.6
17:16The second half — extracting the right context
17:42Entity resolution and the many Mikes
17:58Redundancy cycles and DAG violations
18:37Where that leaves Varick
19:06Every engagement starts with an audit
19:50Closing — slapping a frontier model on top