Every contact-center demo that matters in 2026 opens with the same scene: a human agent mid-call, and beside them a panel that summarizes the conversation, suggests the next step, and drafts the reply. The agent copilot has become the flagship SKU of the seat-license incumbents because it is the fastest way to put generative AI in front of a buyer without answering the harder question of whether a bot can close the call alone. This post is the buyer's version of that demo: what a copilot does once it leaves the keynote, what it costs under the two pricing models the market now runs on, and what Devotel Orbit ships on the same surface today.
1) Why the agent copilot category is burgeoning in 2026
The last eighteen months turned agent assist from a niche feature into a product line. Genesys Copilot and NICE Copilot moved from preview to generally available flagship offerings, Salesforce folded Agentforce assist into Service Cloud seat tiers, and the conversational-AI specialists, Cognigy and Kore.ai, now ship agent-assist panels beside their autonomous agents. Four forces line up behind the parade:
- Model quality crossed the usability line. Transcript summarization and grounded reply drafts are now usually correct on the common cases, which is what a suggestion panel needs. A copilot that is right most of the time and reviewed by a human is shippable; an autonomous agent with the same accuracy is a liability.
- Buyers want a human in the loop. After two years of autonomous-agent pilots, many operations concluded that the safe first deployment of GenAI is next to the agent, not instead of the agent. The copilot keeps the human on the verdict while the model does the retrieval and the drafting.
- Seat vendors found a monetization path. A copilot add-on priced per seat per month is the natural revenue line for vendors whose installed base is licensed by headcount. It sells without re-platforming a single queue.
- The transcript pipeline already exists. Quality scoring, compliance monitoring, and coaching all ride on the same live-transcript infrastructure, so assist is an incremental feature on top of a pipe the vendors already laid.
The result is a category where "copilot" is the headline and the differentiation hides in the details: which knowledge sources ground the drafts, which actions the panel can propose, and whether the bill tracks seats or usage.
2) What an agent copilot actually does in production
Strip the keynote language and a production copilot is four functions:
- Live-transcript summarization. A rolling summary of the conversation so far, so an agent who inherits a transferred call reads thirty seconds instead of listening to six minutes.
- Next-best-action suggestion. The panel watches for the decision point: the fee dispute that qualifies for a waiver, the missed verification step, the retention offer that applies to this plan.
- Knowledge-grounded response templates. Drafted replies assembled from the knowledge base, with the passages they were assembled from attached as citations.
- Action handlers. Buttons that perform the account change the suggestion describes, gated so a human approves every execution.
A transcript-assist exchange, shaped the way the panel renders it:
Live transcript (00:04:12)
Customer: "I moved house last month and the bill still goes to the old
address, and now you want me to pay a late fee?"
Copilot panel
Summary: Customer relocated 12 Sep; service address not updated; late fee
of EUR 14 applied 3 Oct after the direct debit failed.
Next best action: Waive the late fee (within agent authority, under
EUR 20), then update the service address on this call.
Suggested reply (grounded): "I can remove the EUR 14 fee now and update
the billing address while we are on the call. The new address applies
from the next cycle, so the November bill goes to the right place."
Sources: billing-adjustments policy, section 4; KB-118 "service address
change", rev 9.
Actions: [Waive fee EUR 14] [Update service address]
(agent approval required on both)The citation is not decoration; it is the control that makes the draft auditable. A grounded reply should look like this:
Suggested reply: "Hotspot data runs at full speed up to 30 GB per month;
after that it slows to 1.5 Mbps until the cycle resets."
Sources:
1. KB-221 "Hotspot and tethering limits", rev 14, section 2
2. Plan terms, Business Mobile Plus, "Fair usage", p. 6
Status: grounded (both sources retrieved, no conflict)When the panel cannot produce that citation block, the honest behavior is to say so rather than draft from nothing. And the action-handler layer must stay human-in-the-loop: the copilot proposes the waiver, the agent approves it, and the tenant decides which tools exist in the panel at all. A copilot that silently executes account changes is not a copilot; it is an unaudited autonomous agent wearing a smaller hat.
3) The Devotel Orbit agent-assist surface
Devotel Orbit ships the same four functions as a tenant-configured surface inside the dashboard, under the Agents area:
- Grounding-citation panel (
/agents/grounding-citations): every drafted reply carries the knowledge passages it used, in the shape shown above, so a supervisor can audit why the panel said what it said. - Custom tools (
/agents/custom-tools): the action handlers. Each tool is declared by the tenant, proposed by the copilot, and executed only with agent approval. The complete_checkout tool guide is this surface's newest GA marker: a purchase goes from cart to confirmed payment only behind a two-phase consent gate, with the channel rendering and failure vocabulary documented. That is the GA posture here: tools ship with their consent and failure semantics documented. - One shared knowledge base: the customer-facing voice and chat agents and the human-agent assist panel draw on the same KB. A policy revision lands in both at once; there is no second, stale "agent-assist knowledge base" to reconcile.
- Ambient outreach context: the assist panel sees the same customer context the outbound side works from, so when a customer calls three days after an ambient outreach message, the agent knows which message and why (see the ambient outreach explainer).
- Guardrail analytics (
/agents/guardrail-analytics): the measurement layer. It reports how often suggestions are accepted, overridden, or blocked by the guardrails you authored, which is the feedback loop that tells you whether the copilot is helping or humming.
Everything in that list ships GA on Devotel Orbit today, on usage-based pricing, with no copilot seat add-on.
4) Per-seat vs. per-token TCO
The incumbents price the copilot the way they price everything else: per seat per month, as an add-on on top of the platform seat. Indicative list pricing for the copilot add-ons clusters around $55 to $65 per seat per month. The usage-based alternative prices the assist by what it consumes: tokens. The same cost-math style used in our voice-agent TCO breakdown applies here. A representative assist turn costs roughly 8,000 input tokens (transcript window plus retrieved passages) and 1,200 output tokens. At indicative mid-tier model list prices ($1.50 per million input tokens, $6.00 per million output tokens) that is about $0.02 per assist.
The volume math for a 50-seat team whose agents take around 70 assisted conversations a day, 250 working days a year: 875,000 assists, or about $17,500 a year at token cost.
| Line item | Per-seat copilot add-on | Per-token assist (Devotel Orbit) |
|---|---|---|
| Copilot license, 50 seats | 50 x $60/seat/mo x 12 = $36,000/yr | No seat line; assist draws on usage |
| Assisted conversations | "Unlimited" in the brochure, fair-use throttled in practice | 875,000 assists x $0.02 = ~$17,500/yr |
| Seasonal +20 seats for 4 months | +$4,800 per season | Seasonal volume at the same per-assist rate |
| Quiet month (half volume) | Full $3,000/mo either way | Half the token line |
| First-year total, steady state | ~$36,000 and up | ~$17,500 at indicative token prices |
Two structural differences matter more than the totals. First, the per-seat line bills headcount whether or not the panel is used; the per-token line bills conversations, so it tracks the shape of your actual year. Second, the per-seat model makes the copilot a fixed cost you defend in every contract renewal, while the per-token model makes it a variable cost that scales with the business. Verify current quotes against your own assist volume; the point is the shape of the curve, not the printed constants.
5) Decision rules: assist, automate, or both
The copilot vs. autonomous-agent question is not philosophical; it is a containment-rate measurement. Route by intent:
- Automate the intents your autonomous agent contains at a high rate with QA scores you would defend: password resets, balance queries, appointment confirmation. Here the copilot adds latency to a loop the bot already closes (see the containment-rate deep dive for the math).
- Assist the intents with low containment or high stakes: fee disputes, hardship cases, regulated disclosures, escalated emotion. A human makes the call; the copilot compresses the retrieval and the drafting around it.
- Promote gradually. Start an intent in assist, watch the guardrail analytics, and only move it to full automation once the assisted transcripts show the bot would have landed the same outcomes. The assist panel is thus also your safest automation staging area.
The inflection to watch is where containment crosses your quality floor. Below it, spend tokens on human agents with copilots; above it, spend tokens on autonomous agents. A mature 2026 operation runs both at once, on one knowledge base and one quality rubric, which is precisely why scoring every conversation matters: the rubric is what tells you which side of the line each intent sits on.
Frequently asked questions
Is an agent copilot the same as real-time agent assist?
A copilot is the 2026 product packaging of real-time assist: live summary, next-best-action, grounded drafts, and gated action handlers in one panel. The older term usually meant a narrower hint feed; the current category implies the retrieval and grounding layer underneath.
Does Devotel Orbit charge a per-seat copilot fee?
No. The agent-assist surface is priced as usage, like the rest of the platform. There is no copilot seat add-on to negotiate, which is the per-token column in the table above.
Can the copilot apply account changes by itself?
Not on Devotel Orbit. Custom tools are proposed by the copilot and executed only with agent approval, and the tenant chooses which tools exist in the panel. That human-in-the-loop gate is a tenant-owned control, not a vendor default you have to request.
What happens when the copilot cannot ground a reply?
The citation panel reports the missing or conflicting sources instead of drafting from nothing. An ungrounded suggestion is surfaced as ungrounded, which is the difference between an assist and a hallucination with a logo.
When should an intent move from assist to full automation?
When assisted transcripts show the autonomous path would have reached the same outcome at your quality floor: containment above your threshold, guardrail analytics clean, and QA scores you would defend in an audit. Promote one intent at a time, and keep scoring both lanes after the move.