Everyone is buying AI tools. Almost no one is redesigning the work.
Why AI adoption keeps failing in the same places, why an agent is five things at once, and what to do about it.
The pattern is familiar by now. The licences are bought. An AI tool is switched on for everyone. A use case register is set up because the policy says so. Six months later the CEO asks what the benefit was, and nobody can answer. Everyone says we are using it a lot, and sure, a lot of information is thrown around now. Stubbornly, KPIs don’t move.
Leaders are confused and under pressure to obtain a return. Teams are confused about how to make it happen.
Our view is that AI is something entirely new, not just another technology, and there is no established culture and practice for using it. Organisations are running the traditional technology playbook on something that is not a traditional technology. An AI agent has autonomy like a person, procedures like a process, it reads and writes systems like an application, it carries context like data, and it runs on shared models like infrastructure. Five things at once. Most organisations have a separate owner for each one, and none of them owns the agent. Split it across those owners and it fails in predictable ways.
Architects are the people placed to join it back up. We call that role the forward deployed architect.
Watch the session
On 22 October, we shared why AI adoption fails and what to do about it.
The stories come from our own agents and from client engagements across government and regulated enterprise over the past year.
Where it breaks, and why
Before the session we polled our community on why AI initiatives fail to deliver benefits. The answers fall into four patterns. Users notice each one first. The cause sits further up the organisation.
01
Dumb AI
The agent cannot do the job people expected. The vendor answer is to pay more for a smarter model.
Technology drove the requirements. The tender had hundreds of requirements about risk and none about whether the thing would be useful. This is the Copilot trap: a platform chosen for cost and control, then asked to do work it was never tested on.
02
Out-of-the-loop AI
Confident advice that is out of date or misses a policy.
The context was wrong or stale. One advisory agent told a user there was no conflict of interest in a relative auditing their accounts, because the policy it was given predated the rule on perceived conflicts. It only had “approved” data, and approval lagged reality.
03
Over-aligned AI
It refuses basic work and cites a rule it cannot name.
The process owner locked every step down so the agent could never make a mistake. One agent refused a routine change on the grounds of “compliance”. Asked which law, it did not know.
04
AI bureaucracy
Longer documents, more of them, and teams throwing AI output at each other.
Nobody redesigned the flow end to end. A procurement team used AI to write 300 requirements for a cheap purchase. No supplier bid. The review that followed was a 30-page AI-written plan. Months lost, and nobody had asked what they were buying.
Each failure comes from one of the five dimensions being left out. Technology drove the requirements, so the agent is dumb. The data owner approved stale context, so it is out of the loop. The process owner wrote procedures for a person, so it is over-aligned. Nobody owned the end-to-end flow, so the work got slower while everyone got busier.
An agent is five things at once, and nobody owns all five
The cause is deep rooted in how organisations are designed and structure work. Our architecture models are an excellent example: they have a box for people, a box for process, a box for applications, a box for information and a box for technology. An agent sits in all five. Manage it by department and you will be very good at one of the five and terrible at the rest.
| Dimension | What the agent has | Who owns it today | What breaks when they work alone |
|---|---|---|---|
| People | Autonomy. It acts within the authority you give it | HR, the business owner | Nobody sets what it may decide, so it decides everything or nothing |
| Process | Skills baked in. Procedures it follows without being told | Process owner | Procedures written for people get locked in |
| Application | It reads and writes your systems | Application team | It becomes a feature of the platform and the use case bends to the tool |
| Information | Loaded with context about how the organisation works | Data governance | Approved data lags what it needs to know |
| Technology | Runs on shared models and infrastructure | Infrastructure, security | Cost and control choose the model |
There is a second difference that matters as much. An agent does not learn on the job. It works from the skills and context it is given and fails at anything outside them. It is not accountable for the outcome either.
An agent displaces work. It does not displace accountability.
So how do you make it work end to end? It cannot be managed like a person, and it cannot be managed like software or infrastructure. Most organisations have no framework for it yet, which is why the question “who manages the agent?” produces silence in the room.
What to do instead
Guidance for architects.
-
Set the ambition for the organisation
If the goal is to produce more paper, give everyone Copilot and you are done. If the goal is money or customer outcomes, think about the end to end. Decide how far you will push and what risk you will accept.
-
Redesign the work across teams and agents
Map one workflow end to end with the people in it. Decide where an agent runs, where a person decides, where output is reviewed, and what the receiving end can absorb. An agent that sends 400 briefs a day to a team that reads 20 has moved the problem.
-
Manage each agent as one integrated unit
Give it a job description before it gets a prompt. Goals, autonomy and decision authority first. Who it delegates to, when it escalates, and who is accountable when it is wrong.
-
Accept calculated risk
Step-by-step control produces an agent that refuses basic work. Give it room to act and manage the failure modes, because that is where accountability sits.
-
Be ready to adapt
Nobody can agree a target state when the tools change every few months. Agree a six-month horizon and a navigation path for the decisions after it.
-
Learn together
Learning is the human part. The agent cannot do it, so someone has to feed what the team learns back into how the agent is set up.
How we do it
We built an entire Agentic Suite that takes an idea from inception to a full business case, enabling and shaping the decision end to end to reduce waste, increase quality and reduce friction across the many necessary enterprise functions: business users, business partners, business analysts, architects, portfolio managers, strategic planners, financial analysts and procurement specialists.
Initially it was very difficult. Everyone just wanted to reproduce the process and controls, and turn it into monster prompts. The agents became insanely smart and obsequious bureaucrats nobody wanted to deal with.
The model was capable. Its working environment was not. We changed the approach. We redefined jobs with outcomes so each agent could collaborate autonomously: clear goals and limits, the tool set it needed, models matched to the reasoning each step needs, explicit rules for what counts as evidence, and policies reorganised so the agent could find them.
The agent became a mediator, a precious assistant, a smart advisor, cutting weeks into hours, keeping the focus and bringing people on the journey.
Continue the conversation
We design how agents work together across the enterprise: how they work with each other and with your people, what each one may decide, and where the value lands at organisation scale. If you want a second pair of eyes, bring us one workflow and its owner and we will show where the system breaks before it does.