Start with outcomes, not agents. Climb an autonomy ladder and earn each step with evidence.

Risks vs Reward

Business and technology leaders (COO, CIO, CX/Sales/Service directors, Finance partners) who are under pressure to “do something with agents” but are wary of cost, risk and a vendor landscape that changes every quarter.

Gartner predicts that by the end of 2027, 40% of agentic AI projects will be cancelled because of escalating costs, unclear business value or inadequate risk controls (Gartner). Over a short series of posts I will try to answer those failure points in turn, 1. Getting started and maintenance, 2. Cost and Value, 3. The architecture choices that drive best value, 4. Risk controls and release sign-off , 5. Makes sure the value is realised through adoption .

Where to get started

Most organisations I speak to are in the same place. Leadership wants to see agents in the business. Teams have watched the demos. And yet nobody wants to be the person who signs off a pilot that quietly turns into an unpredictable monthly bill, a support burden, and a rebuild every time the vendor ships a new release.

That caution is healthy. Gartner’s own analysts have said that most agentic AI projects are early-stage experiments “mostly driven by hype” and “often misapplied”, and that many use cases positioned as agentic do not actually need an agentic implementation (Gartner press release). The answer is not to wait. It is to start in a way that keeps cost, risk and maintenance visible from day one.

THE LEADER’S QUESTIONS Before we spend a penny… “Which work should I hand to an agent first?”“How much should it do on its own on day one?”“How do I stop the bill and the upkeep running away from me?”

Start with the work, not the technology

The best first agent use cases share five traits:

1. High volume, low variance — the same type of request, many times a week.
2. Information-heavy — the hard part is finding and assembling knowledge, not making a judgement call.
3. A named owner — someone in the business who will own the outcome, not just the tool.
4. A measurable baseline — you know today’s handling time, error rate or first-contact resolution.
5. Low blast radius — if the agent gets it wrong, a human catches it before a customer or a ledger does.

Internal knowledge assistants, case summarisation, proposal first drafts, policy Q&A and meeting follow-up are typical candidates. Autonomous refunds, pricing decisions and anything regulated are not where you start.

Climb the autonomy ladder — and earn every rung

THE ASK “How far should I let it go on its own?”

Think of agents on a ladder rather than as a single category:

Rung	Pattern	Human role	Typical first use
1. Assistive	Agent answers, drafts or summarises when asked	Human decides and acts	Knowledge Q&A, drafting, summarising
2. Semi-autonomous	Agent proposes and prepares actions	Human approves before execution	Case triage, next-best-action, quote preparation
3. Autonomous (bounded)	Agent acts within defined limits and escalates exceptions	Human monitors and handles exceptions	Password resets, order status, appointment changes
4. Autonomous (orchestrating)	Agent coordinates other agents and systems end to end	Human governs by policy and audit	Multi-step fulfilment, cross-system service journeys

Gartner frames it well: the challenge is creating agents “reliable enough to earn greater autonomy”, and organisations that move too quickly to full autonomy risk discovering that “scale amplifies cost, security exposure and unpredictable failures” (Gartner). Each rung up should be a decision backed by test results and production evidence, not an aspiration.

Take the fear out of cost

THE ASK “What if the bill surprises me next month?”

Three practical moves stop cost becoming a surprise:

• Use what you already own first. Many organisations already have AI assistance in their productivity suite and CRM. Exhaust that before building.
• Pick a metered pilot with a ceiling. Agent platforms now offer pay-as-you-go billing — for example, Copilot Studio meters usage at $0.01 per Copilot Credit with no up-front commitment (Microsoft Licensing Guidance), and Agentforce offers no-commitment PayGo Flex Credits (Weflow). Set a budget alert and a cost-per-outcome target before you go live.
• Design for fewer steps. Consumption is driven by what the agent does, not how many people use it. Some vendors set every action to costs the same regardless of tokens, so a 20-step agent costs twenty times a one-step agent . Simple design is cheap design.

Take the fear out of maintenance

THE ASK “What happens when the vendor changes everything next quarter?”

The landscape will keep changing. You cannot stop that, but you can make change cheap:

• Keep business logic out of prompts where you can. Put rules in flows, APIs or reference data that can be versioned and tested.
• Build a golden test set from day one. A few dozen real questions with expected answers lets you re-run quality checks every time a model, a knowledge source or a platform release changes.
• Prefer open integration standards. Model Context Protocol (MCP) for tools and Agent2Agent (A2A) for agent-to-agent hand-offs are now supported across the major platforms, which reduces the cost of swapping components later (covered in Part 3).
• Give every agent an owner, an inventory entry and a retirement date. Agents that nobody owns become the maintenance problem.

A first 90 days

THE ASK “So what does the first quarter actually look like?”

Weeks	Focus	Exit criteria
1–3	Pick 2–3 use cases, capture baselines, name owners	Signed-off value hypothesis per use case
4–8	Build assistive (rung 1) versions on existing platforms	Golden-set pass rate agreed and met
9–12	Run with a controlled user group; measure	Cost-per-outcome and time saved against baseline
Decision	Scale, iterate, climb a rung, or stop	Evidence pack reviewed by business and IT

Stopping is a valid outcome. A cheap, fast “no” is worth more than an expensive, slow “maybe”.