A VP of operations at a mid-market manufacturer told me last quarter that his company “had an AI strategy now.” I asked what it was. He said they’d bought Microsoft 365 Copilot licenses for the leadership team and the project managers. That was the strategy. Procurement had signed off, IT had flipped the switch, and a calendar invite had gone out announcing the rollout.
Six weeks later he called back, frustrated. People were using it to clean up emails and summarize the meetings they’d missed. Nice, but not the transformation the board had been promised. Worse, his best estimator had asked Copilot to pull the margin history on a specific customer across the last three years of jobs, and it returned a confident, fluent, completely useless answer. The data it needed lived in their ERP and an aging estimating tool, neither of which Copilot can see. The estimator now trusts it a little less for everything.
This is the most common pattern I see right now. A company buys Copilot, calls it strategy, and then quietly discovers the gap between a productivity feature and an actual plan for putting AI to work on the problems that matter to the business.
None of this is a knock on Copilot. It’s a genuinely good tool that does specific things well. The problem is the framing. So let’s be precise about where it helps, where it doesn’t, and what a real AI strategy actually looks like.
What Copilot Is Actually Good At
Copilot earns its keep in a narrow but real band: working with the unstructured content already sitting in your Microsoft tenant. Email, documents, chats, calendar, meeting transcripts. When the raw material is text and it lives in Microsoft 365, Copilot is legitimately useful.
The wins are concrete:
- Drafting. First drafts of emails, proposals, status updates, and policy documents. It won’t produce your final version, but it gets you past the blank page, which is often the expensive part.
- Summarizing. Long email threads, dense Word documents, and Teams channels you’ve been ignoring. The summaries are good enough to triage with.
- Meeting recall. This is the sleeper feature. Copilot’s transcription and recap of Teams meetings is the single capability I hear the most unprompted praise for. “What did I commit to in that call?” is a question it answers well, and that genuinely saves people time.
- Search across your own stuff. Finding the deck from two quarters ago when you can’t remember what it was called. Semantic search over your own files beats keyword search most days.
For a knowledge worker who lives in Outlook and Teams, these add up. We’ve seen teams recover a meaningful slice of the low-value administrative hours that pile up around communication and documentation. That’s a real benefit and worth paying for.
But notice the shape of every one of those wins. They all operate on content that is already text, already in Microsoft 365, and already visible to the person asking. That’s the boundary. Cross it, and the value drops fast.
Where It Quietly Disappoints
The disappointment is rarely dramatic. Copilot doesn’t crash or refuse. It produces something plausible, which is exactly what makes the limits dangerous. Here’s where the wheels come off.
It can’t see your business
Your margins, your inventory, your job costing, your CRM pipeline, your production schedule, your CAD models, your historical bid data. The information your business actually runs on lives in line-of-business systems: ERP, MES, PLM, project management tools, custom databases. Copilot has no native access to any of it.
You can connect some of it through Graph connectors and building agents in Copilot Studio, but that’s not a checkbox. That’s an integration project with data modeling, permissions design, and ongoing maintenance. The out-of-the-box product knows your inbox. It does not know your business. When someone asks it a business question anyway, it answers from general knowledge and whatever scraps it found in documents, and it sounds just as confident as when it’s right.
Permissions sprawl becomes everyone’s problem
Here’s the one that keeps security teams up at night. Copilot respects existing Microsoft 365 permissions. That sounds reassuring until you remember the actual state of permissions in most tenants.
For years, oversharing didn’t matter much because discovery was hard. Sensitive files sat in some forgotten SharePoint site or an over-permissioned Teams channel, technically accessible to half the company but effectively invisible because nobody knew to look. Copilot removes that friction. Ask the right question and it will cheerfully surface the salary spreadsheet, the unannounced reorg deck, or the M&A folder that someone shared “just with the team” four years ago.
Copilot didn’t create the exposure. It exposed the exposure. But the practical consequence is that rolling out Copilot safely requires a permissions and data-governance cleanup first, and most organizations have not done one. That work is the project. The license is the easy part.
It doesn’t know your process
Copilot can summarize a document about your change-order process. It cannot execute your change-order process, enforce your approval rules, or flag when a submittal is missing a required signature. It has no model of how your business actually works, because that knowledge lives in people’s heads and in the logic of your systems, not in your documents.
Why Purpose-Built AI Is a Different Animal
This is the distinction that gets lost in the marketing. Copilot is a horizontal productivity layer. It’s the same product for a law firm, a hospital, and an aerospace supplier. Its job is to make individuals modestly faster at generic office work. That’s a fine job. It is not the same thing as putting AI to work on what makes your business specific.
Purpose-built AI starts from the opposite end. It starts with a process that costs you real money or real risk, connects directly to the systems that hold the relevant data, and is built to do one valuable thing reliably. A few examples of the shape, framed illustratively:
- A model that reads incoming RFQ documents, matches them against your historical bid data and current material costs, and drafts a first-pass estimate for a human to review.
- A system that watches your project management and accounting data and flags jobs trending toward margin erosion before they blow the budget, not after.
- A tool that ingests inspection reports and warranty claims to surface the failure patterns your engineers would otherwise need months to notice.
None of those are Copilot features, and none of them ever will be, because they depend on your data, your processes, and your definition of a good outcome. The difference isn’t intelligence. The underlying models are often the same. The difference is integration, context, and a narrow goal someone took the time to define.
There’s a useful way to think about the split. Copilot makes your people a bit faster at the work they already do. Purpose-built AI changes what work has to be done at all. The first is a productivity tool. The second is a strategy. Confusing them is how companies end up with a license and a story instead of a result.
What a Real Strategy Looks Like
A strategy is not a product you buy. It’s a short list of decisions about where AI is worth the effort and where it isn’t. The good news is that the list is usually short, because most of the value hides in a handful of processes.
Here’s the sequence that actually works:
- Start with expensive problems, not available tools. Where do you lose hours, margin, or sleep? Estimating accuracy, scheduling, quality escapes, proposal turnaround. Pick problems with a real dollar figure attached, then ask whether AI can move them. Do not start by asking “what can Copilot do for us.”
- Follow the data. A problem is only a good AI candidate if the data to solve it exists and is reachable. If your bid history lives in three incompatible tools and a senior estimator’s memory, your first project might be consolidating that data, not building a model.
- Fix governance before you scale access. Run the permissions cleanup. Know where your sensitive data lives and who can reach it. This pays off whether or not you ever deploy a single AI feature.
- Buy the horizontal layer, build the vertical capability. Copilot for the generic productivity gains is a reasonable buy. The work that differentiates you, the estimating, the scheduling, the domain judgment, is where custom effort earns its return.
- Measure against the problem, not adoption. “How many people used Copilot this week” is a vanity metric. “Did proposal turnaround drop” is the one that matters.
Done in that order, Copilot has a clear place. It’s the convenient productivity layer riding on top, handling the email and the meeting notes while the purpose-built systems do the work that actually separates you from your competitors. That’s a coherent picture. “We bought licenses” is not.
If you’re staring at a Copilot rollout that hasn’t delivered what the slide deck promised, the honest first step is usually to separate the two conversations: keep the productivity tool for what it’s good at, and start a real one about the two or three expensive processes where custom AI would actually pay off. That second conversation is the strategy, and it’s the one worth the time. If it’s useful to think it through with someone who has built these systems, that’s the kind of thing a 30-minute call or a scoped engagement through Ryshe Forge is for. Either way, the framing matters more than the tool. Don’t let a license stand in for a plan.
