AI Strategyelitics engineering team

AI Agents for Business: From Pilot to Production in 2026

AI agents for business: where they pay off, why most pilots stall, and the checklist, governance, costs and 90-day plan to run one in production safely.

An AI agent for business is software that takes a goal, decides the steps and uses your own systems (CRM, ERP, ticketing, email) to complete a task, with a person approving anything that cannot be undone. Nearly every company has tried one; few run one in production that pays off. The gap is rarely the model. It is a narrow task, an evaluation set of real cases, the right permissions and a named owner who measures the result.

The numbers show how wide that gap is. In McKinsey's State of AI in 2025 survey, 62% of organisations are at least experimenting with AI agents, but only 23% are scaling an agentic system anywhere in the enterprise, and in any single business function no more than 10% report scaling agents. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls.

What an AI agent is for a business (and what it is not)

Three kinds of software get called "AI" in most organisations. A chatbot answers questions. A workflow automation runs the same steps every time. An agent decides which steps to take, calls tools to take them and checks the result before moving on. Each has a different cost, risk profile and way of failing.

Chatbot / assistantWorkflow automationAI agent
What it doesAnswers questions from documents or a knowledge baseRuns a fixed sequence of steps when triggeredPlans steps towards a goal, calls tools, checks results
Who decides the next stepThe userWhoever designed the flowThe model, within limits you set
Touches your systemsUsually read-onlyYes, through predefined connectorsYes, through tools it chooses at run time
Handles exceptionsHands back to the userBreaks or routes to a queueReasons about them or escalates to a person
Best forSelf-service answers, internal knowledgeHigh-volume, predictable processesVaried inputs, multi-step tasks, judgement within rules
Main riskWrong or invented answersBrittle when inputs changeWrong actions, runaway cost, manipulated instructions

A practical rule: if the steps are always the same, build a workflow. It is cheaper and easier to audit. If an answer is all that is needed, a good assistant is enough (see chatbot development). Use an agent where the input varies, the task spans several systems and the next step depends on what the last one found. Many good production agents are mostly workflow, with the model making a few decisions in the middle.

Where AI agents pay off today: use cases and examples

Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025. "Task-specific" is the key word: agents that pay off do one task with a clear input, a clear output and a measurable result.

Use caseWhat the agent doesAccess it needsHuman checkpoint
Support triage and resolutionClassifies tickets, pulls order history, answers known issuesHelpdesk, order system, knowledge baseHandover to staff for refunds, complaints and low-confidence cases
Invoice and document processingExtracts fields, matches invoices to orders, flags mismatchesInbox or document store, ERP (read), supplier recordsFinance approves before anything is booked or paid
Sales ops and CRM updatesLogs calls and emails, updates deal stages, drafts follow-upsCRM, email, calendarRep reviews outbound emails; changes are logged and reversible
IT and internal service deskResets access, answers policy questions, gathers diagnosticsITSM tool, identity provider (limited scopes), internal docsApproval for new access rights and anything touching production
Procurement and order follow-upChases supplier confirmations, compares quotes, updates ordersERP or purchasing system, email, supplier portalBuyer signs off any order, price change or new supplier
Compliance checksChecks contracts or onboarding files against a rule set, lists gapsDocument store, rule set, case managementA compliance officer decides; the agent prepares the file
Research and reportingGathers internal and web data, drafts reports with citationsData warehouse (read), approved web sourcesOwner reviews before the report is shared

The pattern in the last column is the point. In each of these AI agent examples, the agent does the gathering, matching and drafting, and a person decides where money, customers or legal exposure are involved. The agent removes the slow, repetitive part of the work; the checkpoint keeps the risk where it was. Many of these agents also rely on retrieval over internal documents, covered in our guide to advanced RAG.

Why most AI agent pilots stall

McKinsey found that only 39% of respondents report any enterprise-level EBIT impact from AI. Gartner's cancellation forecast names three causes: escalating costs, unclear business value and inadequate risk controls. Gartner also warned of "agent washing", rebranding chatbots and assistants as agents, and estimated that only about 130 of the thousands of vendors claiming agentic AI offer real agentic capabilities.

In practice, stalled pilots share the same handful of problems:

  • No value metric. "Automate support" is not a target. "Resolve order-status tickets without a person, at a quality the support lead signs off" is.
  • No evaluation set. The demo worked on five hand-picked examples. Nobody collected a few hundred real cases with known outcomes, so nobody knows how often the agent is wrong.
  • No access to real systems. The pilot ran on exported spreadsheets. Connecting to the live CRM or ERP, with real permissions and messy data, turns out to be most of the work.
  • Cost surprises. Agents loop. Hard inputs take many more model calls than the test cases did, and nobody set a step or spend limit.
  • Missing risk controls. Security sees the agent a week before launch and, reasonably, asks who approved its access.
  • Agent washing. The "agent" that was bought is a relabelled chatbot that cannot take the actions the business case assumed.

There is also a measurement trap. In the Stack Overflow Developer Survey 2025, 66% of developers named "AI solutions that are almost right, but not quite" as their biggest frustration, and 46% distrust the accuracy of AI tools against 33% who trust it. A 2025 randomised study by METR found experienced developers using early-2025 AI tools took 19% longer while believing they had been about 20% faster. METR has since said those results are out of date, so the lesson is not that AI slows people down; it is that perceived and measured speed can diverge. Measure the agent against a baseline, not against how the demo felt.

Start with a fixed-price agent pilot

We agree the task, the value metric and an evaluation set of your real cases up front, build the pilot on your systems, and end with a clear go/no-go decision. You get a proposal within 24 hours of the first call.

The production checklist for AI agents

A production agent is a demo plus the controls that let you trust it with real work. Before it touches live customers or live money, it should pass every item on this list.

  1. A narrow task with a measurable outcome. One task, one owner, one number, such as resolution rate or days to supplier confirmation. Record the baseline first.
  2. An evaluation set of real cases. A few hundred historical examples with the correct outcome, awkward ones included. Every change to prompt, model or tools runs against it before release.
  3. Tool permissions equal the user's own permissions. The agent acts for someone and can never do more than that person could. No shared admin keys.
  4. Approvals before irreversible actions. Payments, bookings, deletions, customer emails and access changes wait for a human click until the evaluation data says otherwise.
  5. Step and spend limits. Maximum steps and cost per task plus a daily budget; the agent stops and escalates when it hits one.
  6. Logging and tracing of every action. Every model call, tool call, input and output is recorded against the task, so you can replay what happened and why.
  7. Prompt-injection defences. The OWASP Top 10 for LLM Applications lists prompt injection as the top risk (LLM01). Treat every email, document and web page the agent reads as untrusted, separate instructions from data, and require confirmation for sensitive tools whatever the content says.
  8. Data handling and provider settings. Decide which data may reach which model provider, under which contract and retention settings, and where logs live. Mask what the task does not need.
  9. Monitoring and fallbacks to humans. Dashboards for quality, cost per task and escalation rate, alerts when they drift, and a switch that turns the agent off without breaking the process.

AI agent governance and security for the enterprise

Enterprise AI agents fail audits for familiar reasons: nobody owns them, nobody can say what they did, and their access grew quietly. Good AI agent governance is ordinary IT governance applied to a new kind of user.

A named owner

A business owner accountable for results and a technical owner for operations. The owner approves changes to scope and access.

Audit logs

Tamper-resistant records of each action, whom it acted for, what it read and who approved what, with retention agreed with compliance.

Access reviews

The agent's identities and tool scopes join the same periodic review as human accounts. Unused permissions are removed.

Incident handling

A runbook: how to pause the agent, find affected records from the logs, reverse actions and inform customers if needed.

Change control

Prompts, tools and model versions are versioned like code. A model upgrade is a release that runs against the evaluation set first.

Data boundaries

Clear rules on personal and confidential data, provider contracts and regions, alongside the GDPR work you already do.

On regulation: the EU AI Act places extra obligations on AI used in high-risk areas such as credit, employment and education. An agent that matches invoices is in a different position from one that screens loan applications. The Act's timelines have been changing, so check current dates and your classification with counsel before scoping work in a sensitive area. See also our security services and the GDPR outsourcing guide.

Integration: APIs, MCP and why it is most of the build

An agent is only as useful as the systems it can reach. The model call is a few lines of code; connecting safely to your CRM, ERP, helpdesk and document stores is where the engineering time goes. Each integration needs authentication on behalf of the right user, scoped permissions, rate-limit and timeout handling, retries that do not create duplicates, test environments with realistic data, and errors the agent can reason about.

The Model Context Protocol (MCP) helps. It is an open standard introduced by Anthropic in November 2024 for connecting AI applications to tools and data, since widely adopted by major AI providers and tools. Wrapping a system as an MCP server lets different agents and models reuse the same tool definitions instead of rebuilding them each time.

MCP standardises the connection; it does not remove the work behind it. Someone still decides which operations to expose, what each may change, how it is logged and how it behaves when the underlying system fails. Older systems without clean APIs often need an integration layer first, which is where legacy modernisation and agent work meet. Our stack here is typically Python (FastAPI) with LangChain, the OpenAI API or Gemini, and Pinecone or Weaviate where retrieval is needed.

What an AI agent costs

"How much does an AI agent cost?" has two answers, because there are two bills. The build bill is the engineering to scope, integrate, evaluate and harden the agent. The run bill is the monthly cost of operating it: model usage, hosting, monitoring, human review at checkpoints and upkeep of prompts, tools and the evaluation set.

What drives the build bill, roughly in order of impact:

  • Number and state of integrations. A modern SaaS API is quick; a legacy ERP with no API is not.
  • Read or write. An agent that reads and drafts needs far fewer controls than one that books, pays or changes access.
  • Evaluation. Building the case set and the harness to run it is real work, and it tells you whether to go live.
  • Guardrails and governance. Approvals, limits, logging, injection defences and dashboards.
  • Running cost per task. Model choice, number of steps and context size, mostly fixed by design choices in the build.

To make this concrete, here are illustrative ranges at our published rates of €45–65 per engineer-hour, assuming 40-hour weeks and 8-hour days. They are planning examples, not quotes.

Illustrative scopeTeam and durationHoursIllustrative cost
Pilot: one task, one or two systems, read-only or sandbox, evaluation set and go/no-go2 engineers × 3–4 weeks240–320 h€10,800–20,800
Production agent: full path to live, several systems with write access, approvals, limits, logging, monitoring3 engineers × 10–12 weeks1,200–1,440 h€54,000–93,600
Next agent on the same foundations: reuses integrations, logging and evaluation harness2 engineers × 4–6 weeks320–480 h€14,400–31,200
Ongoing care: evaluation runs, prompt and tool updates, model upgrades1 engineer × 2–4 days a month16–32 h per month€720–2,080 per month

For the run bill, track cost per completed task: model and infrastructure cost plus human review time, divided by tasks finished correctly. As an illustration only, at €0.05 of model usage per task, 10,000 tasks a month is €500; whether that is cheap depends on what the same work costs today. Go deeper with our AI development cost guide, get a range for your scope from the cost estimator, or see pricing for rates and squads from €15k a month.

A 90-day plan from pilot to production

Ninety days is a realistic frame for taking one narrow agent from idea to live use on a share of real volume, if the task is well chosen and the systems are reachable. Each phase ends in a decision, so the production budget is only spent on a pilot that earned it.

WeeksFocusWhat happensDecision at the end
1–2FramePick task and owner, agree value metric and baseline, collect the evaluation set, map systems and permissions, early security and data reviewIs the task worth a pilot, and can we reach the systems?
3–6PilotBuild on real systems in read-only or sandbox mode, run the evaluation set, measure quality, cost per task and time saved, run in shadow mode beside the teamGo/no-go: does it beat the baseline at acceptable cost and risk?
7–12ProductionAdd approvals, limits, logging, injection defences and monitoring; access review; release to a slice of volume, widen as the numbers hold; hand over runbooksWiden scope, start the next agent, or hold and improve

A "no" at go/no-go is a result, not a failure: you spent a pilot budget instead of a production budget, and you know why.

Questions to ask an AI agent development partner before signing

With so much agent washing in the market, the partner matters as much as the model. These questions separate teams that build for production from teams that build demos.

  1. How will we measure success, and what is the go/no-go criterion for the pilot?
  2. Who builds the evaluation set, from which real cases, and is it run on every change?
  3. How are tool permissions scoped, and can the agent ever do more than the user it acts for?
  4. Which actions need approval, and is that enforced in code rather than in the prompt?
  5. What are the step and spend limits, and what happens when the agent hits one?
  6. How do you defend against prompt injection from emails, documents and web pages?
  7. Where do our data and logs go, and which model providers are used under which settings?
  8. Fixed price or time and materials, and what exactly is in scope?
  9. Who owns the code, prompts and evaluation set, and does the work happen in our repositories and cloud accounts?

Our answers: AI pilots are fixed price with an agreed evaluation set and a go/no-go decision; NDA and IP assignment are signed before work starts; we work in your repositories and cloud accounts, and you own the code. Engineers start about 48 hours after you approve their profiles, with a two-week trial you don't pay for if you're not satisfied, from Prishtina on CET/CEST, the same hours as Berlin and Paris. More on the AI agent development page; for questions to ask any vendor, see how to vet a software development company.

Frequently asked questions

What is an AI agent for business?

Software that takes a goal, plans the steps and uses your business systems through tools to complete a task, such as resolving a ticket or matching an invoice to an order. Unlike a chatbot it acts; unlike a fixed workflow it decides the next step from what it finds, within permissions, limits and approvals you define.

What are good examples of AI agents in business?

Support triage with handover to staff, invoice and document processing, CRM updates, internal IT service desk requests, supplier and order follow-up, compliance checks against a rule set, and research and reporting. The ones that pay off do one narrow task with a measurable outcome.

How much does an AI agent cost?

There is a build bill and a run bill. Illustratively, at our published rates of €45–65 per engineer-hour, a pilot with 2 engineers for 3–4 weeks comes to €10,800–20,800, and a production agent with 3 engineers for 10–12 weeks to €54,000–93,600. Track the run bill as cost per completed task.

How do you build an AI agent for a business?

Pick one narrow task and a value metric, collect an evaluation set of real cases, connect the agent to the systems it needs with the user's own permissions, and test against the set. Then add approvals, step and spend limits, logging, injection defences and monitoring, and release to a slice of real volume before widening.

How do you keep AI agents secure?

Give the agent no more access than the person it acts for, require approval for irreversible actions, treat every input as untrusted (prompt injection is the top risk in the OWASP Top 10 for LLM Applications), log every action, include agent identities in access reviews, and decide which data may reach which model provider.

Should we build an agent or buy one?

Buy when a product covers your process and systems as they are, after checking that it really takes actions and is not a relabelled chatbot. Build when the value is in your own data, systems and rules. Standards such as MCP make custom integrations reusable across models.

Have a task in mind for an agent?

Tell us the process and the systems involved. Within 24 hours of the first call you get a fixed-price pilot proposal, including the evaluation set and the go/no-go criterion.

Let’s build what comes next.

Tell us what you need to build or scale. You will get a proposal and estimate within 24 hours.

Or write to us directly

hello@elitics.io+383 49 171 069

Dukagjini Center, Prishtina, Kosovo