Customer support automation
Support is the function where AI genuinely earns its cost, because the input is unstructured language arriving in unpredictable shapes. It is also where badly applied AI does the most visible damage. Here is where the line sits.
Automate the sorting, the assembling and the drafting. Do not automate the sending, unless the message is purely transactional. A customer can tell when nobody read their message, and the moment they can tell, the automation has cost you more than it saved.
1. Inbound triage and routing — $10,000 to $16,000
What happens now: a shared inbox two or three people half-watch. Every morning starts with sorting: what is billing, what is technical, what is sales, what is spam. Duplication where two people answer the same thing, and gaps where everyone assumed someone else had picked it up.
What we build: each message classified on arrival by intent, the account reference extracted, priority set from stated urgency and customer tier, and routed to the right queue or owner. Messages scoring below your confidence threshold — the genuinely ambiguous ones — land in a single review queue rather than being guessed at.
Why it costs more than a comparable workflow: roughly a third of the build is assembling an evaluation set from your real historical email and measuring accuracy against it before launch. If nobody can tell you the accuracy number, nobody has measured it. See how we build agents.
2. Drafted replies for approval — $10,000 to $18,000
What happens now: the same forty questions, answered from scratch or from a macro that never quite fits, several times a day.
What we build: a draft written against your own documentation and the customer's actual account state, queued for a person to send with one click or edit in ten seconds. Answers cite the source document, so the agent can check them.
The boundary: drafts are drafted, never sent. If the model cannot find the answer in your documentation, the correct output is "I don't know" and a handoff — not a plausible guess. That distinction is most of the engineering.
3. SLA tracking and escalation — $8,000 to $12,000
What happens now: response commitments tracked by memory and goodwill. Breaches discovered when a customer points them out.
What we build: a clock per ticket driven by your SLA matrix, warnings before breach rather than after, and escalation to a named person on a timer. Reporting on where time is actually going — first response, waiting on customer, waiting on engineering.
Deterministic, deliberately. No AI anywhere in this one. It is pure rules, and rules are cheaper, faster and predictable.
4. Ticket enrichment and context assembly — $8,000 to $13,000
What happens now: before answering anything, the agent opens the CRM, the billing system, the order history and possibly a shipping portal, to find out who this person is and what they bought.
What we build: all of that pulled automatically and attached to the ticket on arrival. Plan, tenure, open orders, recent tickets, payment status, and any flags. The agent starts with the context instead of assembling it.
Underrated. This is frequently the highest hours-saved item in support and almost nobody asks for it, because the assembling feels like part of the job rather than overhead.
5. Feedback collection and routing — $8,000 to $11,000
What happens now: CSAT requests sent inconsistently, results landing in a dashboard nobody opens, and negative responses going nowhere in particular.
What we build: requests triggered on resolution with sensible frequency capping, responses logged against the account, and detractor responses routed immediately to someone who can act — with the ticket context attached.
Review gating — routing satisfied customers to public review sites while unhappy ones go to a private form. It violates the terms of most review platforms, and it makes your own rating meaningless as a signal. Ask everyone, route the negatives to someone who can fix them, and do both.
6. Knowledge base maintenance — $9,000 to $14,000
What happens now: documentation written once and never revisited, while the same questions keep arriving because the answer is wrong or missing.
What we build: clustering of incoming questions to surface what is being asked that the documentation does not cover, flagging of articles contradicted by recent replies, and draft updates prepared for a human to review and publish.
Note the direction: the automation identifies the gap and drafts the fix. A person decides what gets published. Auto-publishing documentation is how you scale wrong answers.
Where the AI line sits, precisely
| Task | Deterministic or AI | Why |
|---|---|---|
| SLA clocks and escalation | Deterministic | Pure rules; a model adds cost and unpredictability for nothing |
| Context assembly | Deterministic | Lookups by account ID; no interpretation required |
| Intent classification | AI | Input is unbounded natural language |
| Detail extraction | AI | Order numbers and dates buried in prose |
| Reply drafting | AI, human sends | Language generation, but consequences need a person |
| Routing after classification | Deterministic | Once you know the intent, the destination is a rule |
Most well-built support automation is a deterministic workflow with one or two model calls inside it. If the design is mostly model calls, something has been over-engineered — see agent or workflow?
What we would do first
Context assembly, then triage, then drafts. Context assembly is deterministic, cheap, immediately useful and carries no accuracy risk, which makes it an ideal first build. Triage comes second, once there is real data to evaluate against. Drafted replies last, because they need both the classification and the documentation to be in good shape.
Drowning in a shared inbox?
Tell us the volume and the systems. We will tell you which of these to do first.