Everyone in this market is converging on the same promise — an agent that resolves instead of deflecting. What separates them is how the action actually happens: through a help article, through your backend, through raw pixels, or through the interface you already built.
Retrieves the right help article and hands it over. Cheap, fast, and it never finishes the job — the hard tickets still reach a human.
Calls your APIs to perform the action. Genuinely resolves — once every system is integrated, which is a programme measured in quarters.
Operates the screen from raw pixels. No integration needed, but it reads a picture rather than the page, and the failures cluster on the actions.
Reads the live page, resolves the actual control, and operates it behind a confirm gate. No integration, no guessing.
| Approach | Completes the action | No integration project | Works apps without APIs | Grounded, not pixels | Per-action audit |
|---|---|---|---|---|---|
| Deflection bots | no | yes | n/a | n/a | n/a |
| Backend agents | yes | per system | no | yes | yes |
| Computer-use | unreliable | yes | yes | pixels | coordinates |
| Inflection | yes | yes | yes | yes | yes |
Approach-level comparison. Individual products vary — see the vendor notes below.
These are strong products built by serious teams. The question isn't whether they work — it's what you have to build before they do.
The best-funded pure play — reportedly around $4.5B valuation on roughly $481M raised, handling a reported 70–75% of conversations end-to-end in its deployments. Resolves through backend integrations: Stripe, Shopify, Salesforce, Zendesk.
Where it stops: every action needs the integration behind it, and pricing reportedly carries a platform fee on top of per-conversation cost. Pick them if you're an enterprise with clean APIs and the team to wire them.
Founded by Bret Taylor and Clay Bavor; reported at roughly $104M ARR growing ~354% year over year, with valuation stepping from about $4.5B to a reported ~$15.8B. Agents call APIs to look up orders, process refunds, change subscriptions.
Where it stops: enterprise-first, with depth achieved by wiring into billing and CRM systems. Pick them if you're a large brand buying a deeply bespoke deployment.
The most accessible outcome-priced agent at a published $0.99 per resolution, aimed at SMB and mid-market, and reportedly being acquired by Salesforce. Sits on top of your helpdesk.
Where it stops: it is excellent at answering from your knowledge base; performing actions in your own product is not the mechanism. Pick them if most of your volume is genuinely informational.
Priced around $1.50–$2.00 per automated resolution. Notably, Zendesk's own CEO has argued publicly that the industry games the word "resolved" by counting deflections and non-answers.
Where it stops: it lives in the ticketing layer, so resolution means resolving a ticket rather than changing something in your product. Pick them if you're standardising on Zendesk anyway.
Around $2.00 per conversation, on top of Service Cloud per-user licensing, with implementation typically a substantial project in its own right.
Where it stops: the total cost and time-to-first-resolution are the highest here. Pick them if your business already runs on Salesforce and the data is there.
Positions as an AI customer concierge for complex, regulated companies — fintechs and healthtechs — resolving end-to-end across phone, SMS, chat, email and WhatsApp, with compliance as a headline feature.
Where it stops: the closest positioning to ours, and the sharpest on regulated verticals. Resolution still runs through the systems you connect. Pick them if you are a regulated business buying breadth of channel and a compliance story today.
Builds a support agent that uses your tools, connected conversationally. Their own framing is the clearest statement of the industry's precondition: if a tool has an API, you can connect it.
Where it stops: that sentence is the wedge. Everything your customer needs that isn't behind a clean API stays out of reach. Pick them if your stack is modern, API-complete, and already documented.
An agentic support platform aimed at B2B, where humans and agents collaborate on customer work — investigating, resolving and acting on signals across the account.
Where it stops: strongest as a B2B support workspace rather than a way to operate your end-customer product. Pick them if your support is account-managed B2B and lives in shared channels.
PolyAI, Parloa, Cognigy, Cresta, Replicant and Kore.ai lead the contact-centre layer, with real enterprise deployments and serious certifications behind them.
Where it stops: these are exceptional at the conversation — turn-taking, latency, barge-in, accents. What happens after the conversation still depends on the integrations you build into core systems. Pick them if voice volume is your problem today.
The only other approach that works your interface without integrations. It processes raw pixels with a virtual mouse and keyboard, deliberately avoiding page and OS APIs.
Where it stops: published results put it well below human performance, and the failures concentrate on exactly the write actions support depends on. See the benchmarks below.
Also in the field. Frontline agents: Ada, Forethought, Crescendo, Gorgias (e-commerce), Fini. Voice infrastructure that these products are often built on, rather than competing with: Retell, Vapi, Bland. And the incumbents bundling agents into existing suites: Salesforce Agentforce, Zendesk, Freshworks Freddy, Genesys, NICE, Five9, Talkdesk.
Read the field end to end and one shape appears. The hard problems this industry has genuinely solved are conversational: understanding intent, holding context, sounding human, handling a phone call without an awkward pause.
Intent, tone, memory, multi-channel, latency, escalation timing. This is genuinely good across the category now.
Whether anything actually changes in your systems still depends on integrations someone has to build and maintain, one at a time.
We work the interface your team already uses, so the action layer doesn't wait on an API programme.
Computer-use is the closest thing to Inflection's approach, so it deserves the most scrutiny. Published benchmarks tell a consistent story: reading a screen is close to solved, and acting on one is not.
OSWorld success for OpenAI's computer-using agent, against 72% for humans on the same tasks.
Best zero-shot result on real Salesforce CRM workflows in the SCUBA benchmark — about 50% even with demonstrations.
Reported end-to-end latency on single desktop tasks, with agents taking 2.7–4.3× more steps than necessary.
One benchmark author put the split plainly: read tasks are largely solved, while write-heavy tasks — logins, forms, submissions — are where the gap opens up. That is the entire job in support. Inflection avoids the problem by not guessing: it resolves a named element through your test ids and accessibility tree, so the action has an identity, a fallback, and an audit record.
| Vendor | Model | Reported price | Plus |
|---|---|---|---|
| Intercom Fin | per resolution | $0.99 | no seat fee with an existing helpdesk |
| Crescendo | per resolution | $1.25 | — |
| Zendesk AI | per automated resolution | $1.50–$2.00 | committed vs. pay-as-you-go |
| Sierra | per resolution | ~$1.50 (estimated, not published) | enterprise contract |
| Decagon | per conversation / resolution | ~$0.99 | reported platform fee ~$50K/yr |
| Ada | per interaction | ~$1.00–$3.50 | — |
| Agentforce | per conversation | $2.00 | Service Cloud seats + implementation |
| Inflection | per resolution | $0.79 | no platform fee · unlimited seats |
Publicly reported figures, point-in-time and subject to change. Not all vendors publish list pricing. See Inflection pricing.
We would rather lose a deal than lose a deployment. Four situations where you should buy something else.
If almost none of your volume needs an action performed, a knowledge-base agent will serve you better and cost less.
Inflection operates an interface. If your customer-facing surface is phone and email with no web or mobile product, there is nothing for it to work.
If procurement requires a completed SOC 2 report today, we are not there yet — and we will tell you that before you spend the time.
Inflection is a resolution layer, not a ticketing system. Keep the helpdesk you have and run Inflection alongside it.
Market and adoption forecasts: Gartner press release, March 2025 (80% of common issues resolved autonomously by 2029); Grand View Research (AI for customer service, $13B in 2024 to ~$84B by 2033).
Deflection-versus-resolution reporting and the Klarna reversal: CNBC, April 2026, including remarks from Zendesk CEO Tom Eggemeier on how "resolved" is counted.
Vendor funding, revenue and positioning: Sacra, TechCrunch, Bloomberg, BusinessWire and vendor help centres.
Pricing: vendor-published list pricing where available; Sierra's per-resolution figure is a third-party estimate and is marked as such.
Computer-use benchmarks: OpenAI's computer-using agent results (OSWorld, WebArena); SCUBA enterprise CRM benchmark, September 2025; OSWorld-Human latency study; Skyvern Web Bench on the read/write split.
Figures are point-in-time in a fast-moving market and may have changed. If you believe something here is out of date or unfair, tell us and we will correct it.
The honest test: bring the ticket your current tool escalates most, and see whether Inflection finishes it in your own app.