The 60 to 70 Percent Automation Target: Where It Comes From and Where It Fails

A COO I was talking with last month pulled up a slide from an industry report on her laptop. Big number on it: "AI will automate 90% of back-office work by 2027." She tapped the screen and said, "So that's what we're aiming at."

I asked her what percent of her back-office work was automated today. Long pause. Then, "Honestly, maybe ten. Probably less."

That gap, from ten to ninety, in under two years, with no diagnosis of what's actually happening in her ops, is the mistake. It's not ambition. It's a category error. The number on that slide is describing something different from what she's trying to do, and chasing it the way she was planning to chase it is how firms end up with a more expensive version of the mess they started with.

The honest target for a mid-market operator between $1M and $50M in revenue is 60 to 70 percent task automation. That's one of three KPIs we track with clients, alongside Away-From-Desk Autonomy and Revenue Per Employee. The number is specific, and people misunderstand it constantly. Let me explain where it comes from and what goes wrong when firms aim higher.

Where the 60 to 70 percent number actually comes from

The number is not a share of all work. It's a share of a specific slice: repeatable, low-judgment work, after the conditions underneath it have been installed.

Break a mid-market ops team's week into three buckets. Bucket one is repeatable, low-judgment work. Invoice matching, status updates, data entry, first drafts of standard emails, simple CRM hygiene, scheduling, routing. Bucket two is repeatable but judgment-heavy. Pricing calls, client escalations, exception handling, anything where the right answer depends on context the team carries in their heads. Bucket three is one-offs and relationships. New deals, hard conversations, strategy, creative work.

For a healthy firm, bucket one is somewhere between 40 and 60 percent of the team's hours. The 60 to 70 percent target is a share of bucket one, not the whole week. Do the arithmetic. If bucket one is half the work and you automate 65 percent of it, about a third of total hours move through the system with approval rather than human doing. That's the scale of change we're after. That's what gets you to Revenue Per Employee numbers that actually move.

The layer four KPI climbs as the first three layers come into place. Context, Data, and Intelligence have to be installed before automation is worth running. Without them, the system is guessing. You can read more about that sequence on the AIOS page. The pattern is almost mechanical: firms that try to automate before they've organized the substrate underneath end up with the same month-four failure we've written about in why AI pilots die in month four.

HBR's writing on automation has been saying a version of this for a decade. The firms pulling real productivity out of automation are not the ones with the most ambitious percentages on the slide. They're the ones who knew what to automate before they picked a tool.

What lives in the other 30 to 40 percent

Once you accept that the target is a slice of a slice, the next question is what lives in the part you don't automate. The answer is the work that actually makes the firm worth hiring.

Judgment work. A client calls, upset, about an invoice. The right response depends on who the client is, what the relationship has been, what else is going on in the engagement, and whether the account manager has a sense that something bigger is wrong. None of that is in a database. It lives in the people who do the work. Automating it produces confident, fast, wrong answers that blow up trust.

Relationship work. The way your senior people talk to clients. The way they read a room on a quarterly call. The way they spot an upsell or a churn risk in the first two minutes of a meeting. That's not repeatable work. It's pattern recognition built on years of context. Automating the shell of it, auto-sent check-ins, auto-generated QBR decks, reliably makes clients feel less cared for, not more.

Edge cases. Every ops team knows its short list of scenarios that break the rules. The billing situation that only comes up twice a year. The one client whose workflow nobody's ever written down. The exception handling that senior staff do on instinct. Building automation for those edges is more expensive than the edge itself. You leave them in the human bucket on purpose.

This is why human-in-the-loop is the default, not the compromise. Approval gates on anything that touches money, clients, or people. Not as a brake. As the design.

Why 90 percent automation is a failure mode, not a stretch goal

The instinct to push past 70 is understandable. You can see the productivity math. Why stop there?

Because past a certain line, the work you're automating is no longer repeatable low-judgment work. It's the judgment, relationship, and edge-case work that should have stayed human, and you've just wired it into a system that handles it badly.

What that looks like in practice: a firm proud of its 88 percent automation number has an AI drafting every client touch, a bot approving expense reports up to some threshold, a scheduler auto-booking meetings based on availability, and a pricing engine quoting standard work. Looks amazing in a deck. On the ground, three things are happening.

One, clients are starting to feel processed. The emails arrive too fast, too polished, too context-free. Senior buyers notice. They escalate less, because they know the next person in line is another bot. The firm's renewal rate starts dipping and nobody can quite tell why.

Two, the team's senior people are spending their time cleaning up automation misfires. They're not saving hours. They're shifting hours from doing work to auditing a system that does work wrong five percent of the time. Five percent of a lot of volume is a lot of cleanup.

Three, the firm has lost its ability to notice exceptions. When humans handled the edges, the team knew the weird cases because they lived them. When automation handles the edges by averaging them into the middle, the weird cases become invisible until they become incidents.

This is why we treat 60 to 70 as a ceiling, not a floor. Past 70 in most mid-market ops, the extra percentage points cost more trust than they save in hours. The work underneath matters. Install in layers, not in leaps. Each layer earns the next.

Why most mid-market ops sit at 5 to 15 percent before diagnosis

Before we start, almost every mid-market firm I meet is somewhere between 5 and 15 percent automated. Some lower.

Not because automation is hard, although it can be. Because they genuinely don't know what's automatable and what isn't. Their repeatable work is tangled into their judgment work. Their client context lives in three places and nobody's reconciled it. Their team can describe their workflows verbally but nobody's written them down. Every "automatable" task has a human exception baked in somewhere that nobody's surfaced.

When a firm in this state buys a tool, what usually happens is they pick one obvious workflow, something like email triage or meeting summaries, and they automate the thin layer of it that doesn't need context. That's a few percent. The tool does its thing. Nobody's life changes much.

Then expansion fails, because the next workflow requires context the tool doesn't have. We wrote about that failure mode in detail in why AI pilots die in month four. It's the same story every time.

The reason most mid-market ops sit at 5 to 15 percent is not that the automation technology is immature. It's that the substrate underneath is scattered. We've written about that cost in the real cost of spreadsheets and Slack and in why more software makes operations worse. Adding tools to a scattered base adds speed, not clarity.

How the number climbs once Context and Data are installed

Here's where it gets interesting. Once Context and Data are installed, the 5 to 15 percent number climbs fast, because you finally know what to automate.

Context means strategy, team, processes, client-handling are organized. You can answer, in writing, what the firm is, who it serves, how it decides, how it communicates. Data means the revenue, ops, and client numbers are centralized and readable by both humans and systems. Not a new CRM. One place where the numbers that matter live.

With those two in place, something changes in what you see. Workflows that felt judgment-heavy reveal themselves as repeatable with a few inputs. Workflows that felt automatable reveal themselves as too entangled with client relationships to hand to a bot. The line between bucket one and bucket two becomes visible, which is what lets you automate one without accidentally automating the other.

Intelligence comes next. Meetings, messages, signals synthesized into briefs your team actually reads. This is where the team starts to feel bandwidth return, before any formal automation kicks in. The McKinsey operations insights hub has good writing on this compounding effect, how intelligence work creates the conditions under which automation is finally worth installing.

Then, and only then, Automate. You can read the sequence on how we work. Task Automation starts climbing because you're automating the right work in the right order. It climbs into the 40s, then the 50s, then the 60s. Somewhere in the 60 to 70 range, it plateaus naturally, because the rest of the work is the judgment, relationship, and edge-case work that should stay human.

The whole time, approval gates are the design, not the exception. The system scores and queues and drafts. A human approves or edits. The loop closes. The system learns from the edits. That's the pattern that earns the climb.

The honest diagnostic question

The question worth asking is not "how do we get to 90 percent." It's this: what percent of our team's repeatable, low-judgment work is moving through a system with approval gates today?

If the answer is under 15 percent, the problem isn't automation ambition. It's that the substrate underneath hasn't been organized. Pushing on tools harder will not move the number. It will probably make your team angrier.

If the answer is in the 40s or 50s and growing, you're on the right path. Keep going. Don't get cute about the ceiling.

If the answer is over 70 and you're proud of it, check your client NPS, your senior team's cleanup hours, and your renewal rate. One of them is probably carrying the cost you're not seeing.

The 60 to 70 percent target is honest because it's bounded by what stays human. It's an engineering target with a human floor under it. Firms that treat it that way get the productivity. Firms that treat it as a starting point on the way to 90 get the cleanup bill.

Our engagement structure is built for this diagnosis. The Fit Check is a free five-minute readiness call. If there's a real fit, we move to a paid Blueprint, the diagnostic that answers where your substrate and your automation ceiling actually are before anyone installs anything. More on what that diagnostic measures in what an AIOS Blueprint measures. If the CEO-as-approval-gate pattern looks familiar, the companion read is the CEO as bottleneck problem.

The right number for your firm depends on your work, not on a slide from an analyst report. Start from where you actually are, install in the order the layers ask for, and let the ceiling show up when it shows up.

-Ed