An AI maturity model for customer service describes how far a support team has moved from human-only service to AI agents that resolve work on their own, and what must be true before it moves further. This model has five levels, from human-led to autonomous processes, six dimensions, and a written gate for every step up.
In short
- Most AI maturity models measure adoption: how much AI you use. This one also measures whether you can prove the AI is right and who controls it.
- The five levels are human-led, assisted, hybrid pilot, scaled hybrid and autonomous processes. Levels 1 and 2 are human support, 3 and 4 are hybrid, 5 is agentic.
- Each level is described on six dimensions: scope, quality measurement, governance, journeys and handoffs, people, and metrics.
- You move up only through a gate: full scoring coverage, quality at or above your human baseline, honored escalations and a named owner who can pause the agent.
- The 15-question self-assessment scores 0 to 45, but your weakest quality, governance or journey answer caps your level, so automation alone cannot buy a higher one.
What is an AI maturity model for customer service?
An AI maturity model for customer service is a staged description of how a support organization uses AI, from none to AI agents running whole processes. A useful one tells you three things: where you are today, what the next level looks like, and the evidence you need before you are allowed to go there.
Most published models stop at the first two. Gartner’s AI maturity model runs from Foundational to Transformational across seven pillars, including strategy, data and governance (Gartner AI Maturity Model toolkit), but it is built for the whole enterprise, not a support floor where an AI agent issues refunds.
The support-specific models we read in November 2026 describe how much work the AI does, rarely how you check that it is right.
| Model (vendor, date) | What it measures | What it leaves out |
|---|---|---|
| Crescendo, Feb 2026: 4 levels | Share of volume automated, how integrated AI is | Quality measurement, owners, approvals; says its levels are “not steps on a roadmap” |
| Puzzel, May 2025: 5 stages | Spread of AI across channels and strategy | Thresholds for moving up; governance only as a trait |
| Observe.AI: 10-question quiz, 6 personas | Data, cloud, integrations, buy-in, culture | Approvals, gates; quality in one question |
| Ada, Apr 2026: operating model | Ownership, operating cadence, shared intelligence | Levels, scoring, gates |
| Cresta, Jan 2025: crawl, walk, run | Rollout approach, pilots | Readiness criteria, governance roles |
That gap matters: a bad bot annoys customers, but an AI agent acting on a wrong policy does it thousands of times, and every case looks resolved. So quality and governance are dimensions you climb here. This is the framework behind our guide to agentic customer service and how to transition: the grid, the gates and a scoring rule.
What are the five levels of AI maturity in customer service?
The five levels are: 1, human-led support with sampled QA; 2, assisted support, where AI drafts and rule-based bots answer FAQs; 3, hybrid pilot, with AI agents on a few narrow intents; 4, scaled hybrid, with AI agents on many intents and actions within limits; 5, autonomous processes run by several agents, with people on exceptions.
| Level | Name | Stage in the transition guide | Model |
|---|---|---|---|
| 1 | Human-led | 1. Human teams | Human |
| 2 | Assisted | 2. Rule-based bots | Human |
| 3 | Hybrid pilot | 3. Single-task AI agents (agentic 1.0), first 2 to 5 intents | Hybrid |
| 4 | Scaled hybrid | 3. Single-task AI agents (agentic 1.0), at scale | Hybrid |
| 5 | Autonomous processes | 4. Autonomous agents running processes (agentic 2.0) | Agentic |
The transition guide uses four stages; this model splits its stage 3 in two, because AI agents on three intents and on forty are different operations: the second needs action tiers, approvals, release control and a dedicated owner.
Two rules keep the levels honest. A level is set per intent (contact reason), not per company: order status can sit at level 4 while complaints stay at level 1. And your level is the lowest dimension you have reached: a team with AI agents on 40 intents but no scoring of their conversations is a level 2 team with level 4 exposure.
The agentic support maturity grid (copy this)
The maturity grid describes each of the five levels on six dimensions: scope of AI, quality measurement, governance, journeys and handoffs, people and roles, and metrics. Read it row by row to find where each dimension sits today: the lowest row sets your level, and the cell to its right is your next piece of work.
| Dimension | 1. Human-led | 2. Assisted | 3. Hybrid pilot | 4. Scaled hybrid | 5. Autonomous processes |
|---|---|---|---|---|---|
| Scope of AI | Nothing customer-facing; macros and help center search | Reply drafts and summaries for agents; FAQ and routing bot | AI agents on 2 to 5 narrow intents; answers and reversible actions | AI agents on many intents; money and account actions within limits | Several agents run multi-step processes; people on exceptions |
| Quality measurement | Small manual sample against a scorecard | Sample of human work; bot containment reports | Every AI agent conversation scored on the people’s scorecard | Every conversation, human and AI, scored per intent and agent version | Quality per journey across agents and people; scores gate releases |
| Governance | Policies in documents and heads | A bot content owner; a flow change log | Owner per AI agent who can pause it; written action limits; AI disclosed | Every action in an autonomy tier; approvals for high-value steps; release and rollback records | Owner per journey; audit trail of every action; approver oversight; rehearsed incident runbook |
| Journeys and handoffs | Transfers between queues, notes in the ticket | Bot hands to a queue; customers repeat themselves | Handover with context; every request for a person honored | Handover quality scored; repeat contact per intent | End-to-end journeys measured, including agent-to-agent handoffs |
| People and roles | Agents, team leads, QA analysts | A knowledge or bot specialist; assist-tool training | A part-time AI agent owner; QA reviews AI conversations too | A dedicated AI agent manager; policy owners; approvers with response targets | Fewer frontline roles; more exception handlers, designers and quality investigators |
| Metrics | CSAT, handle time, sampled QA score | Containment and deflection, added to the above | Correct resolution, critical errors and escalations honored, against the human baseline | Per intent: correct resolution, repeat contact within 7 days, action reversal rate, approval override rate | Journey success rate, cost per correct resolution, incidents per 10,000 conversations, time to detect |
Three rows need a note, because other models skip them.
Quality measurement changes in kind at level 3. A 2% to 5% sample works when people make varied mistakes; an AI agent repeats the same flaw in every matching conversation, so a rare, costly failure hides in the part nobody reads. See AI agent evaluation and one scorecard for human and AI agents.
Governance is the jump from level 3 to 4. OWASP calls the failure “excessive agency” (excessive functionality, permissions or autonomy) and advises to “require a human to approve high-impact actions before they are taken” (OWASP LLM06:2025). Level 4 decides, action by action, which those are: see the human in the loop autonomy matrix and AI agent governance.
Journeys and handoffs decide level 5. When several agents share a process, each can do its step correctly while the journey fails in the handover, so level 5 measures journeys, not agents. See customer journey QA and AI agent orchestration.
What gates must you pass to move up a level?
A gate is a written set of checks an intent must pass, for a sustained period, before it moves to the next level. Write the gates before the rollout, with thresholds signed by CX, risk and IT, so the decision to widen AI is made on evidence rather than on a good demo or a vendor’s roadmap.
The thresholds below are suggested starting points, not industry standards, and match the stage checks in the transition guide.
| Gate | Move | Entry criteria (all must hold) |
|---|---|---|
| A | 1 to 2 | A written scorecard for a good conversation. A human quality baseline per top contact reason. Your top 20 contact reasons listed by volume. A named owner for help center and macro content. |
| B | 2 to 3 | Bot containment checked against real resolution (repeat contact within 7 days). A visible, working path to a person. Contact reasons the bot cannot handle named and routed. Written policies per candidate intent. AI disclosed at the start. A named owner and deputy for the first AI agent. |
| C | 3 to 4 | 100% of the AI agent’s conversations scored. Quality at or above the human baseline on each intent, 4 weeks in a row. Zero unexplained critical errors in 2 weeks. 100% of explicit requests for a person honored. Repeat contact no higher than for human-handled contacts of the same reason. Every action in a written autonomy tier. The pause tested once. A test set run before every change to instructions, knowledge or tools. |
| D | 4 to 5 | Quality measured per journey. An audit trail of every agent action. Approval rules for irreversible or high-value steps, with override rates per approver. A tested way to stop one agent or process without stopping support. The full journey run with synthetic customers and awkward cases before launch. A rehearsed incident runbook. A demotion rule in use. |
Three decision rules settle the arguments gates create.
- An escalation failure blocks the move, whatever the other scores say. A missed handover does more damage than a slow answer.
- Move intents, not the company. If gate C passes for order status and address changes but not refunds, move those two and keep refunds at level 3.
- Gates work both ways. The same critical failure three times in a week drops an intent back a level until the fix is tested and new conversations score clean.
For low-volume intents, add a floor such as the governance guide’s 300 scored conversations in 30 days.
How do you assess your support team’s AI maturity?
Answer the 15 questions below for the intents your AI handles, scoring each from 0 to 3. Add up the points for a total from 0 to 45, read your band, then apply the cap: your weakest quality, governance or journey dimension limits the level you can claim, however much you automate. Score the way things work today, not the way the plan says.
| # | Dimension | Question | 0 points | 1 point | 2 points | 3 points |
|---|---|---|---|---|---|---|
| 1 | Scope | What do AI systems do with customers today? | Nothing customer-facing | Suggest replies, or an FAQ bot | Resolve a few narrow intents | Resolve many intents and act in your systems |
| 2 | Scope | How did you choose the intents AI handles? | Whatever the vendor enabled | Volume only | Volume and policy clarity | Volume, policy clarity and cost of a wrong action, reviewed quarterly |
| 3 | Quality | What share of AI-handled conversations is scored? | None | Spot checks | A regular sample | All of them |
| 4 | Quality | Are AI agents and people held to one standard? | No scorecard | Separate, not comparable | Shared core criteria | One scorecard, comparable per intent |
| 5 | Quality | Do you have a human quality baseline? | No | One overall score | For the top contact reasons | Per contact reason, refreshed monthly, used as the AI’s bar |
| 6 | Governance | Who owns each AI agent? | Nobody named | The vendor or IT | A named owner in CX | Owner and deputy, with tested authority to pause |
| 7 | Governance | Are action limits and approvals written? | No | Informally agreed | Written for money actions | Every action in a tier, enforced, approvals logged |
| 8 | Governance | How are changes to instructions, knowledge and tools controlled? | Anyone edits | Logged after the fact | Logged and reviewed | Run against a test set before release, with rollback |
| 9 | Journeys | What happens when a customer asks for a person? | No path | A hard-to-find path | Usually honored, not measured | Honored every time, and measured |
| 10 | Journeys | Do you measure quality across the whole journey? | No | Per conversation only | Repeat contact per intent | Journeys reviewed across bots, AI agents and people |
| 11 | People | Who works on the AI agents day to day? | Nobody | A team lead, on the side | A part-time owner | Dedicated roles: AI agent manager, knowledge owner, QA for AI |
| 12 | People | How are frontline staff prepared for hybrid work? | Not at all | Tool training only | Trained on handovers and reviewing AI work | Roles redesigned, AI literacy training for all who work with the agents |
| 13 | Metrics | Which number do you report as AI success? | None | Containment or deflection | Resolution rate | Correct resolution rate, next to every volume number |
| 14 | Metrics | Do you track critical errors (wrong actions, invented policies, privacy breaches)? | No | When customers complain | Counted from reviews | Counted from full coverage, each traced to a cause |
| 15 | Metrics | How fast do you detect a new failure pattern? | Unknown | Weeks, via complaints | Days, via reviews | Within a day, via alerts on scored conversations |
Question 12 is also law in the EU: Article 4 of the AI Act requires providers and deployers to take measures to support the AI literacy of their staff and others who operate or use AI systems on their behalf. Since the Digital Omnibus (Regulation (EU) 2026/1744) rewrote it from July 27, 2026, it no longer requires a guaranteed level of literacy (see our EU AI Act guide). Disclosure, part of gate B, comes from Article 50 of the same regulation: AI systems that interact directly with people must be designed so those people are informed they are dealing with an AI system, unless that is obvious.
The scoring rule
- Add the 15 scores for a total from 0 to 45.
- Read the band: 0 to 9 is level 1, 10 to 18 level 2, 19 to 27 level 3, 28 to 36 level 4, 37 to 45 level 5.
- Apply the cap. Average your scores for quality (questions 3 to 5), governance (6 to 8) and journeys (9 and 10). Take the lowest of the three averages. Below 1: you are at most level 2. From 1 to below 2: at most level 3. From 2 to below 2.5: at most level 4. From 2.5 up: no cap.
- Your level is the lower of the band and the cap. If your scope answer (question 1) is ahead of it, you run AI with more autonomy than you can control; close that gap before adding intents. The cap exists because scope is easy to raise and quality and governance are not.
A worked example (illustration, numbers invented to show the math)
A retailer’s AI agent resolves 24,000 of 80,000 monthly conversations (30%) across 12 intents, including refunds up to $50. A QA analyst spot-checks 300 AI conversations a month. The answers: scope 3 and 1; quality 1, 1, 1; governance 2, 2, 1; journeys 2, 1; people 2, 1; metrics 1, 1, 1.
- Total: 21 points, which is the level 3 band.
- Cap: quality averages 1.0, governance 1.67, journeys 1.5. The lowest is 1.0, so the cap is level 3.
- Level: 3. But the scope (question 1) is level 4: a scaled hybrid operation with hybrid pilot controls.
What does that gap cost? The 300 spot checks cover 1.25% of the agent’s conversations. If a knowledge change makes the agent approve refunds outside policy in 0.2% of conversations, that is 48 wrong refunds a month. The sample holds 0.6 of them on average: about a 45% chance of seeing even one, and about 2% of seeing three, enough to call it a pattern (QA sampling statistics shows the math).
The quarter’s plan: score every AI conversation (question 3 from 1 to 3), share the scorecard core (4 to 2), set baselines for the top reasons (5 to 2), review changes (8 to 2), track repeat contact per intent (10 to 2), report correct resolution (13 to 3) and count critical errors from full coverage (14 to 3). That adds 10 points: 31, the level 4 band. The new averages (quality 2.33, governance 2.0, journeys 2.0) lift the cap to level 4, so level and scope match and gate C can be checked per intent.
See it on your own conversations. Kaizo covers the quality measurement row: AutoPilot scores 100% of your AI agent conversations against your own scorecard, the same one your team is scored on, and links every score to the evidence (QA for AI agents). That is the coverage gate C asks for. Book a demo.
How does the maturity model turn into an AI customer service roadmap?
An AI customer service roadmap built on this model is a list of intents, each with its current level, its target level and the gate it must pass next. The order comes from the gates, not a calendar: an intent moves when its evidence holds for the agreed period, usually 4 weeks or more.
- Map every top intent to a level. Expect most at levels 1 and 2.
- Pick the next intents by consequence: high volume, a clear policy, a reversible action. See which intents to automate first.
- Staff the role before the scope. Level 4 needs an AI agent manager who owns instructions, test sets, releases and scores.
- Report correct resolution rate next to containment, so the board sees what the AI got right, not only what it kept.
Choosing a vendor? Turn the gates into AI agent RFP questions: action limits per intent, an export of every action, a pause per intent.
When is a maturity model the wrong tool?
A maturity model is the wrong tool when the work it describes is not yours yet, or when the level becomes a target in itself. Small teams, teams without a scorecard and teams racing to report “level 5” to the board get more from fixing one concrete problem first.
- Small volume. At a few hundred conversations a week, reviewing every AI conversation by hand may be enough.
- No scorecard yet. The cap holds you at level 2; write the scorecard first (gate A).
- Level as a target. Fraud disputes, vulnerable customers and complaints with legal exposure often belong at level 2 for good.
How to start this week
You can place your team on the grid in a week without buying anything: a written starting level per intent, the gap between level and scope, and one gap to close first, with an owner and a date.
- List your top 20 intents by volume and mark each one’s level on the grid.
- Run the 15 questions with the Head of CX, the QA lead and the AI agent owner, separately, then compare: disagreements show where nobody has the facts.
- Apply the cap and write down the gap to your scope.
- Write the next gate for the 2 or 3 intents closest to moving; get CX, risk and IT to sign it.
- Name an owner and deputy for each AI agent, and test the pause once.
Repeat it every quarter; the trend in your weakest dimension tells you more than the total.
Frequently asked questions
What is the difference between an AI maturity model and a CX maturity model?
A CX maturity model measures how customer-centric the whole company is: research, journey design, culture. An AI maturity model for customer service measures how much support work AI does and how well that work is controlled. This model covers only the second.
How long does it take to move from one level to the next?
There is no fixed time, because each move depends on its gate. Gate C needs quality at or above the human baseline for 4 weeks in a row, so an intent cannot move in less than a month. Writing policies and staffing an owner often take longer than the technology.
Can different intents sit at different maturity levels?
Yes, and they should. Order status can run at level 4 while complaints stay at level 1. A company-wide level is only a summary; decisions to widen or narrow AI are made intent by intent, against that intent’s gate.
Will AI agents replace human support agents at level 5?
Not entirely. At level 5 fewer people answer routine contacts, but more work on exceptions, knowledge, agent instructions, approvals and quality investigation. Cutting the team before those roles are staffed is a fast way to lose control of an autonomous process.
What is an AI adoption specialist in customer service?
An AI adoption specialist helps a support team put AI tools into daily work: training agents, changing workflows and tracking use. In this model that work sits in the people dimension at levels 2 and 3, and at level 4 it usually grows into a dedicated AI agent manager role.
