The FDE label is new. The work isn't.

The warning

On 29 September, Gartner predicted that by 2028, 70% of enterprises will abandon agentic AI built by vendor forward deployed engineers, "trapped by soaring costs and unable to evolve it on their own" (The Register). It also warned about "FDE washing", where ordinary consulting gets marketed as something more specialised.

I co-founded and ran a Salesforce consultancy for ten years. Take away the job title and an FDE engagement is what we did every week: discovery, process design, integration, testing, hypercare. So my first reaction was that it's mostly the same work with a better name.

The more interesting part was a quote from Gartner analyst Mukul Saha, who said some providers "charge premium fees without the delivery depth, program management, or change management maturity to justify them."

Why everyone is hiring them

FDE postings on Indeed went from 643 in April 2025 to 5,330 in April 2026 (OneVizion). Between May and July, OpenAI, Anthropic, AWS and Microsoft put roughly US$9 billion into embedding their own engineers with customers. As Forrester noted, "Not one of those dollars went into a better AI model" (Forrester). In September, Accenture and Google Cloud launched a business group that plans to field up to 1,000 FDEs (ITPro).

The money is going into people because getting an agent working inside a real business is the hard bit.

It's a systems integrator project

Look at how the leading AI customer service vendors build. Decagon writes agents as Agent Operating Procedures, natural language instructions that compile into structured workflows. Sierra has "Journeys". Anything that matters, like a refund or a booking change, goes through a tool connected to a real system.

That's the right design, but it means a rollout looks like a software project. Decagon says initial AOP implementation typically takes three to six weeks from kickoff to production (Decagon). AWS sends pods of five or six engineers for around 45 days (OneVizion).

Those engagements follow the same steps as every SI project I ran. Work out what's worth automating, turn the SOPs into something a system can run, connect it to billing and the CRM, test it, then support it after go-live. An Agent Operating Procedure is case routing and escalation rules written in prose instead of Flow.

The washing I'm worried about

Gartner means consultancies relabelling their services to charge more. That's happening, but if the work is the same, I don't find it very alarming.

I'm more worried about the opposite. A vendor sends in a few very good engineers who know the product inside out and ship fast. What they often don't bring is everything SIs built up over years of getting it wrong: a methodology, programme management, sign-off gates, change management, documentation and a handover plan. That's more or less Saha's list. It's the boring part of consulting, and it's what decides whether the thing still works two years later.

Gartner's own advice is to agree governance, IP ownership, knowledge transfer and an exit strategy from day one. That's a decent statement of work. SIs started writing them because customers kept getting burned.

What is different

The vendor employs the implementers. Salesforce pushed implementation out to partners like us to protect its margins. AI vendors are pulling it back in, partly because many charge per resolution. Decagon offers a per-resolution option, a higher rate charged only when the agent resolves a conversation without escalating (Macha). If you only get paid when the agent works, you can't leave it to a mediocre partner.

It's meant to feed the product. Palantir, which coined the term, spent years being called a consultancy before turning what its engineers kept rebuilding into products. Sierra is doing the same with Ghostwriter (Simplify). That's the best argument for FDE over an SI, but Gartner expects that through 2028, fewer than 20% of engagements will turn recurring customer needs into core product capabilities.

Agents don't stay finished. A CRM implementation settles down. An agent drifts as policies change, models get swapped and new edge cases turn up. It's closer to a managed service than a project, and Gartner's 70% is what happens when nobody plans for that.

Most of the work is discovery

I've implemented most of the Salesforce clouds, Agentforce included. On every project, most of the time went on discovery, requirements and documents that played back to the customer what we were going to build. The build was the smaller part. Agentforce was no different.

One example has stuck with me. On a project for a financial services company, we found a consent checkbox with a vague name. Marketing read it as consent to receive marketing. Sales read it as financial services consent. Both meanings had legal weight. The data was clean and complete. It just meant different things depending on who was reading it.

People had worked around that for years because they knew the context. An agent doesn't. Point one at that field and it will pick a meaning and apply it to every customer, and whichever one it picks, someone has a compliance problem. We fixed it by renaming the field and being clear about what it captured. No data quality tool would have caught it. A discovery workshop did.

Agents need more of this than people do. A new support hire learns from the person next to them and fills gaps in a vague SOP with judgement. An agent needs every threshold and exception written down. Decagon's pitch is that you can convert your existing SOPs, which assumes they exist, are up to date and match what people actually do. In my experience that's rare.

AI is good at turning workshop transcripts into a draft requirements doc. It can't decide what happens when two teams handle the same thing differently, or when nobody owns the exception. People used to cover those gaps with judgement. With an agent, someone has to write the rule down and someone else has to agree to it.

Won't AI just build the agents?

Sierra has Ghostwriter and Decagon has AOP Copilot, which drafts procedures from your SOPs (Decagon). They help with writing the flow. They don't help with a billing system that has no write API, two sources of truth for customer identity, or nobody willing to sign off on refunds over a certain amount. Salesforce has sold "clicks, not code" for more than a decade, and SIs are still around.

If you're buying one

Whatever it's called, I'd ask:

• Who runs discovery, and who signs it off?
• What do you own when they leave? The procedures, integrations, tests and IP.
• Who on your side can change the agent after go-live?
• Who looks after it in year two?
• What goes back into the vendor's product? If nothing, you're paying vendor rates for consulting.

Where this goes

If it follows the Salesforce pattern, vendors will do the implementing while the products are immature, then template the common setups (Decagon already ships templates), then hand routine work to partners and keep their own engineers for the big accounts. Accenture's planned 1,000 FDEs look like that handover starting early.

The part I'd keep in-house is the loop back to the product. In the SI model, what a partner learns on a project mostly stays with the partner. Partners find the same gap in the platform on project after project and build the same workaround each time, because the vendor's product team never sees it. An FDE who sits with the customer on Monday and with the product team on Friday can fix it once, in the product. That's the bit an SI can't offer, and it's why I find the model interesting.

I don't know if the FDE label will last. The work won't change much. Before an agent can do a job, someone has to write down everything the job involves, including the field that means two things to two teams. Every enterprise has one of those checkboxes somewhere.

Sources

• Gartner predicts 70% of enterprises will abandon agentic AI built by vendor FDE (Sep 2026) · theregister.com
• Forrester: Forward-deployed engineers are the training wheels for AI reinvention (Jul 2026) · forrester.com
• OneVizion: Forward deployed engineers are a sign enterprise AI has hit the hard part · onevizion.com
• ITPro: Google Cloud and Accenture launch business group to embed FDEs (Sep 2026) · itpro.com
• Decagon: From SOPs to Agent Operating Procedures · decagon.ai
• Decagon: AOP Copilot · decagon.ai
• Decagon: Templates · decagon.ai
• Macha: Decagon AI complete guide · getmacha.com
• Simplify: Sierra company profile (Ghostwriter) · simplify.jobs