The hidden cost of AI agents is the humans behind them

Read the vendors' websites carefully and a pattern appears. Sierra writes: "We staff a dedicated agent engineer and product manager on each of our deployments," and describes weekly meetings with customer stakeholders that never end. Decagon's marketing says its agents configure "without engineering or vendor bottlenecks," while its job board hires an Agent Deployment Engineering organisation in five cities, with a posting that says the quiet part plainly: "It's not enough for our customers to just 'set it and forget it'... Truly successful AI Agents require hands-on execution, rigorous iteration, and deep technical delivery."

Salesforce went further and productized the role: a certified "FDE Partner Network," trained by Salesforce engineering, for "complex integrations, and production-ready Agentforce deployments." Even Intercom, whose Fin promises to be live "in less than an hour," runs a deployment team that "works directly with customers" and holds office hours with solutions architects every two weeks. The hour covers answering from your help centre. The integrations are where the humans come in.

The role has a name: forward-deployed engineer. Palantir invented it, and for years had more forward-deployed engineers than software engineers. In 2025 it became the hottest job title in AI. Postings grew more than 800% between January and September. Salesforce committed to a team of one thousand. OpenAI and Anthropic both embed theirs with their largest customers; Anthropic's job ad promises candidates 25 to 50 percent travel, to customer sites, to build in person.

Venture capital has noticed, and approves. a16z titled its essay on the subject "Trading Margin for Moat: Why the Forward Deployed Engineer Is the Hottest Job in Startups" and argued it is "shortsighted to be optimizing for 80% gross margin." Read that from the customer's side: the services are the strategy, and the margin they trade away comes back on your invoice. It is a model only enterprises can afford. Everyone smaller gets a chatbot and a knowledge base.

If shipping engineers worked, the numbers would look better than they do. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, and names the reasons: escalating costs, unclear business value, inadequate risk controls, and integration with legacy systems that is "technically complex, often disrupting workflows and requiring costly modifications." S&P Global measured it: the share of companies abandoning most of their AI initiatives before production jumped from 17% to 42% in a single year.

MIT's State of AI in Business report found that 95% of organizations are getting zero measurable return on generative AI, and that only 5% of embedded, task-specific AI tools ever reach production. Its diagnosis is ours: pilots stall on "integration complexity and lack of fit with existing workflows." Enterprises take nine months or longer to get from pilot to production. McKinsey's survey of nearly 1,500 organizations found that of 25 attributes tested, workflow redesign, not model choice, has the biggest effect on whether AI shows up in the bottom line.

One forward-deployed engineer summarized the whole field in two sentences: "The model is usually the cleanest part. The hard part is finding the workflow nobody documented."

Gartner estimates that of the thousands of vendors now selling "agentic AI," only about 130 are real; the rest are what it calls agent washing. The thesis explains why the washing works: when the integration is done by humans behind the curtain, the demo and the product can be very different things, and nobody outside the deployment can tell. An agent that integrates itself has nowhere to hide. Either it wires into your site in minutes or it does not.

Integration is not a milestone, because websites do not hold still. Prices move, endpoints change, pages get redesigned. That is why Sierra meets its customers weekly and why implementation orgs keep growing: the work never finishes, it only renews. Any agent whose wiring is done by hand inherits a maintenance contract by construction.

Look closely at what a forward-deployed engineer actually does on an agent project and there are exactly two jobs. The first is discovery and integration: finding the systems that hold the answers and wiring the agent into them. The second is translation: turning what the business wants into how the agent behaves.

We built Algomo to do both itself. It finds the endpoints your pages already call and builds its own tools to use them; that is self-integration. You state rules, limits and behaviour in plain language and it implements them; that is AI-first. The two pillars of the product are the two halves of the job everyone else staffs.

We can make that bet because of where the agent runs. Everyone else integrates server-side, which means API contracts, credentials and meetings: humans. Algomo works inside the visitor's own session, riding the interface your website already exposes to every visitor. The integration surface already exists, and it already works. When your site changes, keeping up is the agent's job, not a change request.

Bold claims need boundaries, so here is ours. Systems behind a login, like your order database or your CRM, are connected together with you: a setup session, not an engineering project. Refunds, payments and anything else you name stay with people, because you say so in plain language and the limits hold. What your pages already do in the open, the agent wires up in minutes.

Every claim above about other vendors is from their own websites, job boards and press, and every number is from the named source. But the claim about us does not need a bibliography. Point Algomo at your site on the free plan and watch it wire itself in. If the thesis is wrong, it will be wrong in front of you, in about three minutes.