
Agent assist gets treated as the timid option — the thing you do if you are not confident enough to ship a customer-facing assistant.
That framing misses what it actually is. Agent assist is the experiment that tells you which intents are safe to automate, and it produces labelled training data as a by-product of doing useful work.
Skipping it does not mean moving faster. It means making the same bet without the evidence.
The sequence
0. contact-driver analysis ← fix what should not exist
1. corpus repair ← the actual blocker
2. agent assist ← proving ground, 1 quarter
3. narrow autonomy ← only intents step 2 proved
4. proactive outbound ← eliminate, don't handle
5. voice ← only if 1–4 are solid
Most programmes start at step 3, discover step 1 was never done, and produce a system that says wrong things fluently.
Step 0: fix what should not exist
Before automating anything, cluster last quarter's contacts by root cause rather than stated intent.
Not "delivery status enquiry" but "shipment delayed past the promised window and nobody told the customer." Not "billing question" but "the invoice line description does not match what they bought."
Frequently the top three clusters are product defects generating contacts as a side effect. Automating those makes them cheap, which makes them invisible, which makes them permanent — the pain that would have driven a fix gets anaesthetised.
A cluster of identical contacts originating from one page in an onboarding flow is not a support problem. It is a missing sentence.
Step 1: the corpus is the blocker
Teams spend evaluation effort on model selection. In deployments I have reviewed, the binding constraint is almost always whether the system can find correct information.
Typical findings in an unaudited corpus: two articles answering the same question differently because one was updated and the other forgotten; policies that changed while documentation did not; answers correct for one market with no metadata distinguishing them; content written for internal readers being served to customers.
Retrieval from that produces fluent, confident, wrong answers — harder to catch than keyword-search failures, because the error now has good grammar.
Minimum metadata to make retrieval a filter rather than a hope:
market: [UK, IE]
product: [premium, plus]
effective_from: 2024-06-01
effective_to: null # null = current
audience: customer
owner: billing-team
last_verified: 2026-07-14
effective_to alone removes a large share of wrong answers, by letting superseded content stay archived without competing for retrieval.
Step 2: agent assist as instrumentation

Here is the part that justifies the whole sequence.
While a human handles a contact, the system retrieves relevant context, drafts a candidate response, and pre-populates the summary and disposition. The human edits and sends. Nothing reaches a customer unreviewed.
The value is immediate — reduced handle time, faster onboarding for new representatives, better consistency because everyone sees the same retrieved guidance.
But the durable value is the data. Every accepted, edited or rejected suggestion is a labelled example. Instrument it:
{
"conversation_id": "c_8f2a…",
"intent": "request_refund",
"suggestion_id": "s_44b1…",
"retrieved_doc_ids": ["kb_112", "kb_884"],
"action": "edited", // accepted | edited | rejected
"edit_distance": 0.34,
"final_text_hash": "…",
"model_version": "assist-v3.1"
}
After a quarter you can answer, per intent, questions you would otherwise be guessing at:
-
What proportion of suggestions were accepted unchanged?
-
Where edits occurred, were they stylistic or factual?
-
Which documents appear most often in rejected suggestions? (That is your content backlog, ranked by impact.)
-
Which intents have high acceptance and low edit distance?
That last one is your automation shortlist, derived from evidence rather than from which intents seemed simple in a workshop.
Step 3: narrow autonomy
Promote only intents that cleared the bar in step 2 — high acceptance, low factual edit rate, and an observable downstream outcome you can verify.
Non-negotiable at this step:
-
One-step escalation, always available, no negotiation. Systems requiring three failures before offering a human generate the complaints that end up public.
-
Complete context transfer on handoff — transcript, detected intent, attempted actions, verified identity.
-
Proactive escalation on frustration, repetition, or high-stakes intent. Cancellations, complaints, bereavements: route immediately regardless of capability.
-
Resolution measured, not containment. Repeat-contact rate within seven days for automated conversations, against a matched sample that reached a human. If automated contacts repeat more, containment is measuring failure.
Steps 4 and 5
Proactive outbound uses the step 0 analysis: detect the predictable condition — delayed shipment, failed payment, expiring card — and contact the customer first. Straightforward automation with a model composing only the message. It eliminates contacts rather than handling them more cheaply, and it is the highest-return pattern in the whole sequence.
Voice last, and only on well-understood intents. Latency is the binding constraint — conversational turn-taking tolerates a fraction of the delay chat does, which rules out multi-step retrieval chains without restructuring. Interruption handling matters more than answer quality for perceived competence. And the failure mode is worse: a frustrated caller has no transcript and is one step from a public review.
The cost note
Model inference cost per turn, not per contact, with a realistic distribution of conversation lengths. A retrieval step injecting several thousand tokens into every turn multiplies per-contact cost quietly; a summarisation and quality-review pass multiplies it again.
Design routing in from the start — a small cheap model for high-volume simple intents, escalating to a larger one for hard cases. Retrofitting a routing layer into a single-model system is awkward, and it is the main lever for keeping the curve flat.
Full guide — retrieval quality, escalation design, personalisation, governance and measurement: AI Customer Experience. We also structure these deployments.
Frequently Asked Questions
Why start with agent assist rather than a chatbot?
Because it produces labelled evidence of which intents are safe to automate while delivering value, at a risk profile where a bad retrieval becomes a rejected suggestion rather than a wrong answer sent to a customer.
What should agent assist instrument?
Per suggestion: the intent, retrieved document IDs, whether it was accepted, edited or rejected, edit distance, and model version. That gives you an automation shortlist and a content backlog ranked by impact.
How long should the proving period run?
About a quarter. Long enough to cover intent variety and seasonal patterns, and to accumulate enough examples per intent that acceptance and edit rates are meaningful rather than anecdotal.
What must be in place before customer-facing autonomy?
One-step escalation always available, complete context transfer on handoff, proactive escalation on frustration or high-stakes intent, and resolution measurement rather than containment.
Why is voice last?
Latency constrains architecture severely — conversational turn-taking tolerates far less delay than chat, ruling out multi-step retrieval without restructuring. Interruption handling drives perceived competence, and a frustrated caller has no transcript and escalates publicly.
How do we control inference cost?
Model per turn rather than per contact with realistic conversation lengths, and design routing in from the start so cheap models handle high-volume simple intents. Retrofitting routing into a single-model system is awkward and it is the main cost lever.