The 4-Gap Framework is how FreshStack diagnoses why AI isn't compounding inside a business. It comes out of assessments run on real teams, and it exists because the same four failures keep turning up regardless of size or sector.
The premise is that AI rarely fails for the reason people assume. Teams reach for a better model or another tool when the actual constraint sits upstream of both: no shared way of working, no connection to where the work lives, no decision about data and accounts, or nobody owning any of it once it's built.
Each gap has a symptom you feel and a name you can act on. The symptom is what shows up in the week. The name is what you fix.
Standardisation: why two people get different results from AI
Standardisation · you feel it as Consistency
Standardisation is whether your team has one shared way of doing a task with AI, or as many ways as there are people. You feel it as consistency: the same job comes back different depending who ran it. Without a canonical version, quality is a personality trait rather than a property of the business.
Left alone, it stops being a quality problem and becomes a staffing one. Someone senior starts checking everything before it goes out, and that review step turns into the new bottleneck. At that point you're paying for AI twice: once for the tool, and once for the person cleaning up after it.
Read: One good loaf is not a bakery →Last-Mile Integration: why AI never has your context
Last-Mile Integration · you feel it as Connection
Last-Mile Integration is whether AI is connected to where your work actually lives, or whether someone hand-feeds it context every time. You feel it as connection: generic output, and the same background pasted in before every task. The last mile is the expensive one, and it's the one most teams skip.
The tax is invisible because it's spread across everybody. Ten minutes of context-setting on every task never shows up on a budget line, but it's the reason AI saves less time than the demo promised. It also caps what you can automate at all: nothing runs unattended if a person has to supply the background first.
Read: AI automations for small and medium businesses in 2026 →Governance: knowing where your AI work actually happens
Governance · you feel it as Exposure
Governance is whether you can say what your AI tools have access to, on whose accounts, under what rules. You feel it as exposure: an uneasy answer when a client asks. It's the gap that stops being yours the moment it becomes an incident, a contract clause or a regulator's question.
This is the only one of the four with a deadline attached to it. Client security reviews and the EU AI Act's transparency duties ask the same question, and “we've never really decided” isn't an answer either of them accepts. It's also the cheapest gap to close early and by far the most expensive to close after something has already gone wrong.
Read: EU AI Act enforcement 2026 →Maintenance: why AI setups quietly stop working
Maintenance · you feel it as Ownership
Maintenance is whether anyone owns keeping your AI setup alive after it's built. You feel it as ownership: prompts go stale, permissions drift, seats go unused and the team slips back to doing it by hand. Most AI that “didn't work” worked fine on day one and nobody owned noticing when it stopped.
Decay is quiet, so it usually gets mistaken for a failed experiment. The tool takes the blame, the project gets shelved, and the next attempt starts from scratch with a more sceptical team. It's why the second AI project inside a business is often harder to sell than the first one was.
Read: How much does AI automation cost →// the order matters
The gaps aren't independent, and the order you close them in changes what AI automation costs.
Governance goes first whenever it's live. Connecting AI to more of your data before you've decided what it's allowed to touch turns a fix into an exposure event, which is why the quiz warns about that pairing specifically.
Standardisation and Maintenance travel together. A shared workflow library with nobody owning it drifts back into freestyling within months, so building one without naming someone to keep it is work you'll end up doing twice.
Last-Mile Integration usually carries the biggest payoff, and usually belongs after the other three are at least decided. Connecting a mess faster only gets you a faster mess.
The 4-Gap Framework is FreshStack's diagnostic for why AI isn't compounding inside a business. The Bottleneck Finder is its 10-question version; the AI Assessment is the full instrument.