of organisations have reached the stage where AI agents run multi-step workflows
That figure comes from a Wakefield Research survey of 1,000 senior technology and data leaders, published by Teradata in July 2026. Yet nine in ten planned to increase their agentic AI investment over the next year.
Because we design, build and then run agents in production across a portfolio of clients, we get an unusually clear view of where these programmes succeed and fail, and what it takes to land agents in production and make them stick. These are the seven mistakes we see most often.
The seven mistakes:
- Handing out licences and calling it a strategy
- Automating workflow by workflow
- Approving pilots with no route to production
- Treating governance as an afterthought, or as a brake
- Assuming the data and systems are ready
- Measuring hours saved instead of outcomes
- Keeping technology and operations apart
1. Handing out licences and calling it a strategy
The most common thing we hear in a first conversation is some version of “we’re already using AI”. What that usually means is that the team has ChatGPT, Copilot or Claude, and people have been left to find their own uses for it.
Those tools were built as personal assistants, and they are good at it. Individuals get quicker. But the business does not change: same headcount, same margin, same manual processes running exactly as before. The gains sit with the individual, so when that person leaves, the gains leave with them. And because nobody has set rules on what goes into these tools, client and company data often ends up somewhere no one has approved.
There is a real difference between your people using AI to make their jobs easier and your business running differently because of it. The first is a procurement exercise. The second is a transformation programme, and in our experience it is the only one that shows up in the numbers. Your competitors bought the same licences at roughly the same time, so the licences alone are not an advantage.
The better question: if every AI licence were cancelled tomorrow, what in the way this business runs would actually change? If the answer is that things would simply slow down, AI has not yet changed how the business runs.
2. Automating workflow by workflow
The most common starting point we see is to pick a workflow, automate it, then pick the next one. It feels low-risk and it produces quick demonstrations. What it rarely produces is a better business, because each automation simply adds speed to a process that was never designed for agents.
Most technology estates evolved by accident: a CRM chosen here, a finance system there, an IT supplier somewhere else, none of them joined up. Automating on top of that locks the fragmentation in. The businesses that see real returns step back first and redesign the operating model: which work agents should do, which work people should do, and how the systems underneath need to connect.
The better question: if we were designing this business today, knowing what agents can do, what would it look like, and what is the roadmap from here to there?
3. Approving pilots with no route to production
Pilots are comfortable. They are small, reversible and easy to celebrate. They are also where most agent programmes quietly die, because nobody agreed what success looked like before the pilot started, or what would happen if it succeeded.
A pilot without a pre-agreed decision gate is not an experiment. It is a demonstration with a budget. Building one has never been easier. Landing it in production, supporting it, maintaining it and getting people to use it is where the real work sits, and it is why so many businesses are stuck in pilot purgatory.
The better approach: make every pilot a step on the operating model roadmap, not a standalone experiment. Agree the success criteria, the decision date and the route to production before it begins. If it passes, it scales. If it fails, stop, and keep what you learned.
4. Treating governance as an afterthought, or as a brake
We see businesses get governance wrong in one of two directions. Some bolt it on after go-live, then discover an agent has write access to the finance system and nobody can explain its last fifty actions. Others treat governance as a reason to say no, and block anything that touches live data.
Both miss the point. Governance is what allows you to give an agent more responsibility over time, safely. A practical way to start is to tier agents by what they can touch:
- Read-only agents that summarise, research and recommend carry the least risk.
- Agents that write into live systems need logging, limits and clear escalation routes.
- Agents in regulated or business-critical systems need the tightest controls and human sign-off on consequential actions.
Whatever the tier, every agent should have a named owner, a full audit trail and a way to switch it off. Consequential actions should still end with a person who reviews and signs them off: the agent takes on the work, not the accountability.
The better question: whose name is on the off switch, and could we reconstruct what the agent did last Tuesday?
5. Assuming the data and systems are ready
An agent is only as good as what it can reach. If customer records live in three systems that disagree, the agent will struggle to reconcile them. It will act unpredictably, and it will do so at speed.
This is the barrier the Teradata survey found most often. 77% of leaders said a fifth or less of their data was described well enough for agents to use reliably.
Integration is invisible in a demonstration, which is why businesses routinely underestimate it. In practice, connecting the agent reliably to the systems it depends on is usually where most of the delivery effort goes.
The better question: if a capable new hire had only the access this agent will have, could they do the job well?
6. Measuring hours saved instead of outcomes
“The agent saves 30 hours a week” is one of the most common success measures we hear, and one of the least useful. Hours saved only matter if they are redeployed into something valuable. Otherwise they disappear into the working week and never reach the P&L.
Measure the outcome the agent exists to move instead: quotes issued, cases closed, days to invoice, pipeline generated or error rate. Record the baseline before work starts, then judge the investment against the value of that outcome rather than against headcount.
The better question: which number will move, by how much, and by when?
7. Keeping technology and operations apart
Many businesses hand AI to IT, or to one enthusiastic manager, or hire a single transformation lead and expect them to carry it. Either way, the technology sits on one side of the business and the operations it is meant to change sit on the other. Once agents take on real work, those become the same thing, and they need owning from the top, with a sponsor at board level.
The technology is also the smaller part of the job. The rest is enablement, change control and culture: how deeply the new ways of working are built into the business, and whether the people who do the work today help design them. They hold the unwritten rules, workarounds and exceptions that no process map captures. Leave them out and the agent will be built on a version of the process that does not exist.
The better question: who in the leadership team owns this, and does the plan cover people and ways of working as well as technology?
What we see in the programmes that work
The businesses that get agents right do not necessarily understand the technology in depth. What they have in common is that they insist on five conditions before approving anything:
- An operating model redesigned for agents, with a roadmap whose work streams compound rather than compete.
- A baseline taken at the outset, with every agent measured against the outcome it exists to move.
- Pilots that sit inside the roadmap, with success criteria and a route to production agreed up front.
- Access tiered by risk, with a named owner and an audit trail for every agent.
- A partner that stays accountable in production, and a system the business owns outright.
The Teradata numbers are a warning about decision-making, not about technology. They also point to a gap that compounds. Once a business stops treating AI as something its staff use and builds it into how the business runs, each improvement funds the next, and the distance grows every month.
This is the work Digital Planning does: AI-native operating model design, followed by the build, governance and day-to-day running that make it stick. We are not a consultancy that hands over a slide deck and leaves, a development house paid by the hour, or an IT provider selling licences. Because we run what we recommend, it is often in our interest to advise what not to build. To talk it through, get in touch.
Sources
- Teradata: New Research: Why Enterprise Agentic AI Stalls Before It Scales (press release, 7 July 2026; Wakefield Research survey of 1,000 senior technology and data leaders, fieldwork March to April 2026)
