A slide that says 'become a Frontier Firm' is a company-sized goal, and nobody does company-sized work on a Tuesday. Here's how the journey looks when you run it one workflow at a time.
You can’t become a Frontier Firm in one move, but you can move one workflow at a time. Place each workflow on Microsoft’s three phases, fix the flow before you hand it over, set the human-agent ratio role by role, and use a five-question test before anything moves up. Run it as a 90-day quarter with a register a leader can read in five minutes.
Say you’re the COO of a 600-person distribution company. Your CEO comes back from a conference with one slide, and the slide says “Become a Frontier Firm by year end.” On Thursday the head of operations asks the only question that matters in the room: “Which part of my Tuesday changes?” Nobody can answer her.
Nothing is wrong with the slide. The trouble is its size. “Become a Frontier Firm” is a company-sized goal, and people don’t do company-sized work on a Tuesday. They work through invoices, onboarding requests, month-end commentary and support escalations, and each of those is a workflow that can move one step at a time or stay where it is.
That’s the argument of this guide. A company doesn’t become a Frontier Firm in one move; its workflows earn each phase, one at a time, on evidence that a named person signs for. What follows is how to run that: a three-phase path for each workflow, redesign before delegation, a test for moving up, a role-by-role split between people and agents, and a quarter you can start this month.
First, where the phrase comes from. “Frontier Firm” is Microsoft’s term, from its 2025 Work Trend Index, which describes organizations built around on-demand intelligence and hybrid teams of people and agents, a blueprint it calls “AI-operated but human-led.”[1] Microsoft sells the Copilot and agent tools that make this possible, so read its surveys as a vendor’s research. The same survey found that 81% of leaders expect agents to be moderately or extensively part of their company’s AI strategy within 12 to 18 months, which goes some way to explaining why the slide reached your desk.
It’s an expectation, not a result. Microsoft’s 2026 edition also applies “Frontier” to groups of surveyed people, such as its Frontier Professionals, as well as to firms, so check which meaning a figure is using before you quote it.[6]
If you want the picture of what changes day to day once people direct agents, our piece on the AI-powered company describes it. This one is about the order in which you get there.
Microsoft describes the journey in three phases. In the first, AI acts as an assistant that removes drudgery. In the second, agents join teams as “digital colleagues,” taking on specific tasks at human direction. In the third, humans set direction for agents that run entire processes, checking in as needed.[1] Its own CEO guide gives the advice I’d put on the wall: “Choose one process in an area of low risk and strong potential, and try making it agent-led.”[2]
Notice what that advice applies to. It’s a process, not a company. The phases are often presented as stages a firm passes through, but in practice a firm sits in all three at once. Finance might use an assistant to draft variance commentary (phase one) while accounts payable (AP) has an agent preparing exception notes for a person to approve (phase two), and I’d keep phase three well away from anything that touches payroll.
| Phase (Microsoft’s wording) | What it looks like in this workflow | Who decides what |
|---|---|---|
| 1. Human with assistant | The AP specialist asks an assistant to summarise a vendor’s email thread and draft a reply. The specialist still finds the exception, matches the invoice and decides. | The person does the work; AI speeds up pieces of it. |
| 2. Human-led agents | An agent reads each new exception, drafts a note and proposes a fix. The AP lead approves every note before anything posts. | The agent prepares; a named person approves each item. |
| 3. Human-led, agent-operated | The agent resolves routine exceptions below an agreed amount and passes the rest up. The lead reviews a weekly sample, plus everything it escalated. | The agent acts inside written limits; the person owns the limits and the exceptions. |
Illustrative example. Phase names follow Microsoft’s Work Trend Index; placing a single workflow on them is our framing. It deliberately differs from the company-level scoring in our AI-ready enterprise framework, which assesses a whole organization rather than one workflow.
A company doesn’t have a phase. A workflow does.
Two things follow. In our view phase two is where most of the learning happens, because it’s where a person finds out what the agent gets wrong while the person is still approving every item. A workflow that jumps from one to three skips the stage in which that judgment gets built.
And phase three is a choice for each workflow, not a destination for the company. Plenty of workflows will sensibly stay at phase one or two for good, because the cost of a miss is high or the volume is low.
Say you lead accounts payable at that distributor. You have an agent that can read exception emails and draft a note, and go-live is Monday. On Friday afternoon you notice that exceptions reach your team by three routes: a shared inbox, a tab called “Stuck” in a spreadsheet, and purchasing messaging you directly. The agent can see one of them.
Put it on the inbox and it makes roughly a third of the work faster while the rest stays invisible to it, and your team now has two ways of doing the same job. Microsoft’s own playbook for its transformation says the pattern plainly: “dropping an AI agent into the existing process often just automates the dysfunction.” In the same document it adds that “a tool licensed and rolled out to 100,000 employees does not change how the work gets done.”[3]
The budget version of this problem is in our look at why enterprises struggle with AI adoption despite record investment. The workflow version starts with three questions, asked of every step before an agent gets near it.
Redesign the work, then delegate it. An agent on a messy process makes the mess faster.
Once the flow is clean, write down who does what. This is the version for the invoice workflow at phase two.
| What AI does | What the human still owns | How it gets checked |
|---|---|---|
| Reads each exception in the single queue and drafts a note with a proposed fix | Approving or rewriting each note, and deciding whether to pay against a disputed invoice | The AP lead initials every note before it posts; the controller spot-checks ten a week |
| Matches the invoice to the purchase order and flags mismatches | Deciding whether a mismatch is a vendor error or an internal one | A mismatch log, reviewed on Fridays, with a one-line reason for every wrong flag |
| Drafts the reply to the vendor | Sending it, and the tone with suppliers who matter most | Nothing goes out without the lead pressing send |
| Nothing, for disputes involving fraud, legal or a supplier in default | Whether to engage at all | The controller and legal sign off in writing |
Illustrative split for one workflow at phase two. Keep the last row even in a small company.
Workflows tend to move up a phase because someone is impatient or the demo went well, and much less often because of evidence. The fix is to write the criteria before the pilot starts, so nobody is tempted to write them afterwards to fit what happened.
| # | Question | What a yes looks like |
|---|---|---|
| 1 | Is there a named owner who signs the move? | One person, not a team, with their name on the register. |
| 2 | Is “done” written down, and have recent outputs been checked against it? | A one-sentence definition and a sample of outputs (20 is a reasonable start) scored against it by someone other than the person who built the workflow. |
| 3 | Is the exception path written? | A list of what the agent must pass to a person, and how quickly a person has to respond. |
| 4 | Is there a way back? | A named person can switch the agent off or reverse its actions, and the team has tried it once. |
| 5 | Has it run a full cycle at the current phase without a serious miss? | A month-end, a campaign or a hiring round, not a good week. |
Scoring: five yes answers, move up one phase for a trial of one cycle. Four, fix the missing one first. Three or fewer, the workflow stays where it is.
The two rows people tend to leave blank are the written exception path and the way back. Both feel like admin until the first time an agent passes along something it shouldn’t have, or acts on something it misread. For how deep the review should go at each level of consequence, our guide to writing human-in-the-loop guidelines covers it, so I won’t repeat it here.
A workflow moves up only when a named person signs for it and a way back is written down.
The same list works in reverse. Decide in advance what sends a workflow back a phase:
Moving down isn’t a failure, just the workflow telling you its owner is stretched or its inputs changed, and a leader who treats a demotion as news rather than blame will hear about problems sooner.
Ask a finance controller, a support lead and an HR business partner how many agents they could supervise properly and you’d probably get three different answers, each right for that role. Microsoft calls this the human-agent ratio and describes it as a new metric organizations will need for human-agent teams.[1] In plain terms, it asks how many agents a role can use and how many people it takes to guide them. The ratio comes from how much one person can review before their review turns into a rubber stamp, rather than from what the software can do.
That’s the failure to watch for as agent numbers rise. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, and the three reasons it gives are escalating costs, unclear business value and inadequate risk controls.[5] It’s an analyst’s forecast, not a count, and none of the three reasons is about how clever the model is. They’re about whether someone owns the value and the controls, which is what the tables above are for.
So set the ratio per role, starting from what each person can genuinely check.
| Role | A good phase-two candidate | What stays human | The ratio question to ask |
|---|---|---|---|
| Finance controller | First drafts of month-end variance commentary | What a variance means, and anything going to auditors | How many drafts can I review properly before close? |
| Operations director | Vendor invoice exceptions, delay notices to customers | Supplier disputes and relationships | How many exceptions an hour can my lead check? |
| HR business partner | First-draft answers to routine policy questions | Anything about a named employee, pay or a leave case | Who reads the escalations the same day? |
| Support lead | Draft replies to first-line tickets | Refunds over a set limit, upset customers, safety issues | What share of replies do I sample, and how fast? |
| Sales operations manager | CRM clean-up and account research briefs | Pricing and discount decisions | Who owns data quality when an agent updates records? |
Illustrative starting points, not benchmarks. The right ratio is the one at which the reviewer is still reading.
If you’re managing the agents yourself, our guide to being an agent boss covers job descriptions and onboarding for them. It’s written for marketing teams, but the habits carry over to any function.
For a leader, the first quarter is less a launch than one workflow that has been redesigned and run at phase two long enough to earn a verdict, a second one being prepared, and a record that shows both. Here’s a version that fits twelve weeks.
Choose a low-risk, high-volume workflow. Answer the three redesign questions, name an owner and a reviewer, and write the “done” sentence. Note how long the work takes today.
The agent prepares, a person approves each item. The reviewer keeps the mismatch log and the owner checks a sample every week.
Score it with someone who didn’t build it. Write down what’s missing rather than arguing about it.
Move up for one trial cycle, hold, or stop. Start the second workflow and show the register to the leadership team.
A starting plan. Lengthen it for workflows that run on a monthly or quarterly cycle.
The record is a register a leader can read in five minutes. It also gives you a place to measure what matters. Logins and message counts tell you whether people are using the thing, which is the first and easiest signal. We use a three-step lens: use, then persistence, then impact. Use is whether people are doing it. Persistence is whether they’re still doing it a month later without being reminded. Impact is whether the work itself is faster, better or different.
| Workflow | Owner | Phase today | Use | Persistence | Impact | Next step |
|---|---|---|---|---|---|---|
| Vendor invoice exceptions | AP lead | 2 | Nearly all exceptions go through the single queue | Samples still reviewed in week 8 without a reminder | Median time to resolve is down; target set in week 1 | Run the test in week 9; decide in week 12 |
| Month-end variance commentary | Controller | 1 | Three of four analysts use the assistant | To be checked after the next close | Fewer review comments on drafts | Redesign first, then trial at phase two |
| First-line support replies | Support lead | 2 | Agents draft most replies | Sampling dropped in week 6 | Escalations unchanged | Hold; restart the sampling before any move |
Illustrative entries. Review the register on the first Monday of each month, and update the phase only on the owner’s signature.
A leader who only sees the Use column is looking at access. The register earns its place in the Persistence and Impact columns, where the third row above shows a workflow quietly losing its review habit before any result has moved.
There’s some outside support for working this way, with a caveat. In McKinsey’s 2026 survey, nearly three-quarters of the respondents it classes as AI high performers said they had fundamentally redesigned workflows because of AI, against about a quarter of everyone else.[4] The answers are self-reported and show a link, not a cause, so it supports the order of work in this guide rather than proving it.
It helps to keep four things apart, because they get used as if they meant the same. Training builds capability: people can direct an agent and check its work, in principle. Workflow design, which is the redesign in section three, gives that capability somewhere to be used. Reinforcement turns it into behavior, through the sample reviews, the monthly register and a manager who asks about both. Adoption is what you see when a workflow keeps running at its phase without anyone chasing it. It’s the outcome of the other three, not a lever on its own.
That’s the same idea as Tool x Workflows x Behavior in our guide to what makes an AI-powered professional, applied to a workflow instead of a person.
Teams can start this internally with the tables above, and many will. Expert-led training shortens the learning curve in specific ways. It lets the AP lead and the controller practise phase-two review on real samples before the pilot, it builds the same review standard across a team rather than in one enthusiastic person, and it covers the judgment calls a guide can only point to, like what to do with an exception that doesn’t fit the list. That’s the work we do in our corporate training sessions.
For this week, list ten recurring workflows in your area. Circle one that’s low-risk and happens often. Write down its owner and its “done” sentence, then put it on the register at phase one or two, whichever is true today.
No. It’s Microsoft’s label, introduced in its 2025 Work Trend Index as a description of where it expects organizations to head, and there’s no certification or independent audit behind it. If a vendor or consultant offers to score you against it, ask what the scoring is based on and who checked it.
A maturity model scores the whole company, usually on a ladder with several stages. This approach places each workflow on its own, so one company can have a workflow at phase three and another at phase one without contradiction. If you also want the company-level view, our AI-ready enterprise framework covers it.
It can, but phase two is where the person approving each item learns what the agent gets wrong, so skipping it removes the stage that builds that judgment. A workflow that’s very low-stakes may move quickly, and for anything touching money, people or customers we’d keep at least one full cycle at phase two.
Each workflow needs a named business owner, because that person decides what ‘done’ means and signs the move up. IT owns the platform and the access controls, HR and the manager own role and workload changes, and someone senior owns the register. One person owning all of it can leave the workflow owners without real authority.
Yes, scaled down. You don’t need a register with three owners, only one workflow, one named person, a written ‘done’ sentence and a way to switch it off. The five-question test works with a team of two, and a small company can move a workflow up a phase faster because fewer people have to agree.
The three phases, the ‘AI-operated but human-led’ wording, the human-agent ratio and the 81% figure come from Microsoft’s 2025 Work Trend Index (31,000 workers in 31 markets, February to March 2025, fielded by Edelman Data x Intelligence). The one-process advice is from Jared Spataro’s CEO guide, and the two playbook quotations are from Microsoft’s Becoming a Frontier Firm playbook, second edition. Microsoft sells the tools it writes about, so treat these as a vendor’s claims. The McKinsey and Gartner figures are from the publishers’ own pages, read in October 2026. We left out secondary-coverage claims about Microsoft’s internal cycle-time and evaluation-trust numbers because we couldn’t find them in the playbook itself. The companies, people and figures in the worked examples are illustrative, not real.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.