If you spend any time on LinkedIn or X right now, you have probably seen the term forward-deployed engineer (FDE). It is being called the hottest role in technology, with compensation running from $150,000 base all the way up to seven figures for the best people. Most of the coverage is aimed at engineers who want the job. This article is for the other side of the table: business owners and operators who want to know what forward-deployed AI actually means, why it matters, and how to get its benefits inside the tools you already own, without hiring a million-dollar engineer.
What is forward-deployed AI?
The term comes from Palantir, which built its business on forward-deployed engineers: technical people who go on site with a customer, learn how the work actually happens, and then customize software around those workflows instead of forcing the customer to adapt to a generic product. The model worked because the value was never in the platform alone. It was in the person who bridged the gap between the business and the technology.
Applied to AI, forward-deployed engineering is the practice of embedding with a team, mapping real workflows step by step, deciding where AI belongs (and where it does not), and then deploying it inside the systems the business already runs on: the CRM, the ERP, the inbox, the ticketing queue.
In plain terms: we embed with your team, map how work actually happens, identify where AI can create measurable value, and deploy it inside HubSpot and the systems you already use. That is forward-deployed AI, and it is what a growing number of businesses are discovering they need before any AI tool will pay for itself.
Buying AI is not a competitive advantage anymore
Here is the uncomfortable truth driving all of this: every company can now buy the same intelligence. New frontier models ship constantly, and any business with a credit card has access to the same foundational capability as its largest competitor. Talk to fifty different companies and you will find them using largely the same AI stack.
When everyone has the same intelligence on tap, intelligence stops being the moat. The advantage moves to deployment: where, how, and why you apply that intelligence to your specific business. Every company is different in how its processes run, what its exceptions look like, and what actually matters to revenue. The businesses that win will be the ones with the best bridge between their own processes and the intelligence they have access to.
Why most AI pilots fail
MIT research made headlines with the finding that roughly 95% of generative AI pilots fail to deliver measurable results. Having watched a lot of AI projects up close, we think the reason is simple: most companies skipped the deployment discipline and went straight to what practitioners call token maxing. Give everyone a license, point the model at everything, and let it figure things out.
There are real horror stories here, including executives who have burned through eight-figure AI budgets in a single quarter with little to show for it, because access was handed out everywhere with no plan for where the intelligence actually belonged.
The fix is not a better model. The fix is judgment about where AI belongs, and that judgment has to start with something unglamorous: understanding how the work is really done.
The documented process is rarely the real process
Ask someone to describe the first step of their workflow and they will say something clean like "an email arrives." The reality is that it arrives from forty different senders, no two formatted alike. The data is in a PDF, or a screenshot, or buried in a forwarded thread. Half the cases are exceptions, and the rules for handling them live in one person's head, not in any SOP.
If you build automation for the documented process, you build for a system that does not exist. This is why forward-deployed AI starts with workflow discovery: sitting with the people who do the work, tracing what actually happens, and capturing the exceptions. There is only one way for a process to go right and a thousand ways for it to go wrong. Automation that only handles the happy path is worth very little. Automation built around the real exceptions is where the value is.
Where AI belongs, and where it does not
The most important judgment call in any AI project is separating three kinds of work:
- Deterministic steps. Predictable, rule-based work: validation, routing, record updates, notifications. This should be conventional automation. If-this-then-that logic and API calls are cheaper, faster, and more reliable than a model. In HubSpot, this is standard workflow automation, and much of a typical process should stay right here.
- Judgment steps. Interpreting intent, categorizing something messy, summarizing context, drafting a response. This is where a model earns its keep. In a ten-step workflow, often only two or three steps truly need judgment.
- Human decisions. Anywhere consequences are high or uncertainty is real, a person approves before anything ships. The AI drafts, the human decides.
Getting this split right is the difference between an AI system your team trusts and one they quietly stop using.
The Selworthy framework: Discover, Evaluate, Deploy, Improve
Selworthy has always started engagements the same way: audit the stack, build a blueprint, then build and scale. Forward-deployed AI gives a sharper name to the loop we run when the work involves AI:
- Discover. Map how the work actually happens: the steps, the systems, the exceptions, the bottlenecks, and what each one costs. The output is an operating map of your workflows with a prioritized list of where AI can create measurable value and where it should stay out.
- Evaluate. Before anything touches production, we build evaluation sets from your real historical data. If the system categorizes emails, we test it against emails you have already handled and scored. An evaluation report might show 41 of 50 runs passing, with the nine failures traced to missing data or a wrong record. That evidence drives fixes before launch, not after.
- Deploy. The system goes live inside your existing stack, starting in shadow mode where it drafts but does not act, then earning autonomy step by step. Every action is logged in an audit trail. If you cannot see what an agent did and why, you should not trust it, and neither should we.
- Improve. Live performance feeds back into the evaluation sets. Measurement stays focused on the only three buckets that matter to a business: revenue uplift, cost savings, and risk mitigation. And once one workflow improves, the next bottleneck becomes obvious, so the loop runs again.
What this looks like inside HubSpot
Here are five workflows we see constantly, split the way a forward-deployed engineer would split them:
| Workflow | Deterministic layer | AI judgment layer | Human control |
|---|---|---|---|
| Lead qualification | Property validation, routing, notifications | Interpret intent and score fit | Sales reviews uncertain leads |
| Service intake | Ticket creation, SLA assignment, record updates | Categorize the issue and summarize context | Agent approves sensitive responses |
| Pipeline hygiene | Required-field checks, stale-deal workflows | Identify risk signals and recommend next actions | Manager approves major stage changes |
| Sales follow-up | Enrollment rules, task creation | Draft account-specific outreach | Rep reviews before sending |
| Data cleanup | Formatting, deduplication, validation | Interpret ambiguous company or contact data | Ops reviews low-confidence matches |
Notice the pattern. Conventional software handles the predictable steps, AI handles the judgment-heavy steps, and people stay in control wherever consequences or uncertainty demand it.
Build on what you already own
One more principle we hold firmly: forward-deployed AI should build on top of your existing systems, not force a migration. If a company just spent two years and serious money settling into its ERP or CRM, an AI proposal that starts with "first, switch platforms" deserves the rejection it will get. The higher-value move is making the stack you already own smarter: connecting HubSpot to your other systems, layering agents and automation on top, and letting your team keep the tools they know.
That is also why we do this work inside HubSpot rather than around it. Breeze agents, workflow automation, custom-coded actions, and API integrations give us everything needed to deploy AI where your revenue data already lives, with the audit trails and permissioning a business actually requires.
You do not need a million-dollar hire
The market is paying enormous salaries for forward-deployed engineers because the combination is rare: consulting-grade communication plus real technical judgment. Most small and mid-market businesses will never hire one, and they should not have to. This is exactly what an embedded agency partner is for. Selworthy has been doing the audit-first, build-on-your-stack version of this work for years; forward-deployed AI is the clearest name yet for how we approach the AI side of it. Credit where it is due: the forward-deployed engineering framework discussed here was laid out in Greg Isenberg's interview with Vas of Varick Agents, and our take builds on that conversation.
How to start: run a sprint, not a science project
Every one of our AI engagements starts with a short, structured discovery effort we call the AI Workflow Opportunity Sprint. Clients consistently tell providers of this kind of work that the audit was worth many times what they paid for it, and also that nobody enjoys the word audit. So: a sprint. It produces a current-state workflow map, a bottleneck and exception inventory, an AI-versus-automation decision matrix, prioritized use cases with expected revenue, cost, and risk impact, human-approval requirements, and a fixed implementation roadmap with a recommended proof of concept.
If you want to see where AI actually belongs in your business, that is the place to begin. Claim your 4 complimentary consulting hours and we will use them to start mapping your highest-value AI opportunities inside HubSpot.
