Last year, nearly every operations leader we spoke with ran an AI pilot. A chatbot on the help desk. A copilot inside the ERP. A proof of concept that summarized shipping exceptions or drafted patient reminders. The demos were impressive. Twelve months later, most of those pilots are still pilots.
When we dig into why, the answer is almost never the model. Today’s models can read an invoice, triage an email, or reason through a scheduling conflict better than most people expected even two years ago. The pilots stalled because the AI had nowhere to do the work.
That is the problem we exist to solve. Brem builds software so AI can work for your business. This post explains what that means, and why it is the difference between a demo and a team member.
The Three Walls Every Pilot Hits
Across the 40+ operations teams we have worked with, stalled pilots run into the same three walls.
Wall 1: The AI can read, but it can’t write
A pilot that summarizes your exceptions is useful. A pilot that summarizes your exceptions and then fixes them is transformative. The gap between those two is almost always access.
Most operational SaaS tools were designed for humans clicking through screens, not software taking actions. APIs are read-only, rate-limited, or gated behind an enterprise tier. So the AI produces a recommendation, a person copies it into another system, and the time savings evaporate.
In one regional 3PL we worked with, an AI model correctly extracted data from 96% of bills of lading in testing. In production, it saved almost nothing, because a coordinator still had to re-key every extracted field into the WMS. The model was fine. The system it needed to write to would not let it.
Wall 2: Nobody trusts it with anything that matters
The second wall is trust. Operations leaders are rightly cautious about letting software take actions that affect customers, patients, or money. When a pilot cannot explain what it did, cannot be limited to certain actions, and cannot be rolled back, the only safe answer is to keep a person in the loop for everything.
That is not automation. That is an expensive suggestion box.
Trust is an engineering problem, not a culture problem. It requires:
- Permissions: the agent can do exactly what its job requires and nothing else
- Thresholds: below a confidence level, a person decides
- Audit trails: every action records what the agent saw, what it did, and why
- Reversibility: anything the agent does can be undone
Very few off-the-shelf AI tools provide all four, and none of them can provide it across your specific systems.
Wall 3: The data is scattered and stale
The third wall is data. An agent that monitors shipments needs live carrier updates, current inventory, and order history in one place. An agent that fills appointment cancellations needs the real-time schedule, the waitlist, and patient preferences.
In most businesses, that data lives in four to eight systems, synced overnight if at all. An AI working from yesterday’s export makes yesterday’s decisions.
What AI Actually Needs to Work
If the model is not the bottleneck, what is? In our experience, AI needs four things from the software around it before it can take on real work:
| What AI needs | What it looks like in practice |
|---|---|
| A place to act | Full read and write APIs into the system of record, not screen-scraping or CSV drops |
| Guardrails | Role-based permissions, approval thresholds, hard limits, and an off switch per agent |
| A memory | Clean, connected operational data available in real time, not overnight |
| A record | An audit log of every action, readable by your team and your auditors |
Sometimes your existing systems already provide most of this, and we build agents that connect directly to them. Sometimes they do not, and the right move is to build the system of record your operation and your agents both run on. Either way, the software comes first. The agent is only as capable as the environment it works in.
Autonomy Is Earned, Not Switched On
The other reason pilots stall is that they ask for too much trust too fast. A team is shown a demo and asked to let the AI run a workflow end-to-end. Unsurprisingly, they say no.
We deploy every agent in three stages:
- Shadow. The agent runs alongside your team on real work and proposes what it would do. Nothing changes in your systems. You measure accuracy on your own data.
- Assist. The agent drafts the entry, message, or action, and a person approves it with one click. This is usually where teams first feel the time savings, and every approval becomes training signal.
- Act. Routine cases run on their own within the limits you set. Exceptions still go to a person, with full context.
Each stage has a clear exit criterion. For an intake agent, we typically look for 95%+ field-level accuracy over two weeks of shadow mode before moving to assist. Teams that follow this path reach full autonomy on routine work in about six weeks, and, more importantly, they trust it.
How to Restart a Stalled Pilot
If you have a pilot sitting on a shelf, here is how we would approach restarting it:
- Pick one workflow, not a platform. Choose the highest-volume, lowest-judgment task your team does every day. Invoice entry, appointment confirmation, and lead response are common winners.
- Map the systems it touches. For each system, ask: can software write to it? Can we limit what it does? Can we see what it did? Every “no” is the real project.
- Measure the baseline. Hours per week, turnaround time, error rate. If you cannot measure the before, you cannot prove the after.
- Run in shadow mode first. Two weeks on real data will tell you more than six months of demos.
- Decide what the agent is never allowed to do. Write down the hard limits before you write a single prompt.
The Bottom Line
The businesses getting real value from AI in 2026 are not the ones with the best models. Everyone has access to the same models. They are the ones that built the software layer that lets AI do real work: a place to act, guardrails, memory, and a record.
That is what we build. If you want to find the first workflow worth handing to an agent, book a 30-minute agent assessment. You will leave with three workflows ranked by hours saved and payback, whether or not you work with us.