Three different groups of people type “Ada AI” into a search bar, and they want three different things. Some are programmers checking whether the search engine has confused a chatbot with the Ada programming language. Some are support directors trying to work out if they can cut queue times without setting the brand on fire. The rest already talk to Ada several times a year without knowing it, usually while chasing a late delivery at 11pm.
The company is Ada, a Toronto outfit founded in 2016 by Mike Murchison and David Hariri. It raised a $130M Series C in 2021 at a reported $1.2 billion valuation, and its agents now handle conversations for Meta, Canva, Zoom, Discord and Shopify. That roster tells you more than any product page: Ada sells to companies with millions of repetitive support contacts and a reputation worth protecting.
What Ada AI actually is
Ada sells one thing, and it’s narrower than the marketing suggests: automated customer service that finishes the job rather than pointing at a help article. The platform connects to your help centre, order management system and CRM, then answers questions across web chat, email, SMS, WhatsApp and voice. A human gets involved only when the AI decides it’s out of its depth.
The distinction Ada pushes hardest is between a bot that answers and a bot that resolves. A deflection bot replies with “here’s an article that might help.” An Ada agent looks up your order, sees the parcel was split into two shipments, pulls tracking on the second box, tells you it lands Thursday, and closes the ticket.
How a resolution actually unfolds
Underneath, there are four steps. The agent reads the message and works out what the person wants. It pulls the relevant facts from your knowledge base and connected systems. It takes an action, whether that’s a refund, a password reset or a delivery reschedule. Then it confirms the outcome in plain language.
The reasoning layer
Older chatbots matched keywords to canned responses, which is why so many fell apart the moment a customer typed something unexpected. Ada’s current AI Agent uses a reasoning model that can plan multi-step actions. Ask “why was I charged twice this month?” and it can pull two invoices, spot the duplicate, explain the cause, and offer a credit in a single exchange. When a request needs information from three different systems, that planning ability is the difference between useful and useless.
Resolution rate is not the same as deflection rate
This is where a lot of vendor reporting gets slippery. Deflection counts anyone who didn’t reply again, which includes people who gave up and phoned your call centre. Resolution counts a query that was actually answered and closed. A deployment can post a 70% deflection rate while leaving customers furious.
Ada reports automated resolution rates in the 60s and 70s for well-tuned accounts. Treat those as the optimistic end of the range, achieved on high-volume intents with clean knowledge bases. A messy help centre will land closer to 30% no matter whose logo is on the software.
Building the agent, and keeping it on a leash
Ada’s AI Agent Studio is where support teams define tasks, connect systems and set boundaries. You can stage changes, run test conversations against recorded transcripts, and roll back a prompt that starts misbehaving. None of that is glamorous, but it’s the difference between a pilot that survives contact with real customers and one that gets switched off in week three.
Guardrails matter more than intelligence
The hard part isn’t answering questions. It’s refusing them correctly. Give an agent the authority to issue refunds and you have to decide the ceiling: automatic up to $50, human approval above it, no exceptions. You also want a complete action log, because “the bot said it was fine” is not a defence in a chargeback dispute.
This is a live issue across the industry. Microsoft’s new AI code of conduct asks models not to trick humans or act outside their remit, and customer-facing agents sit squarely in that territory. If a support bot can move money, someone in legal should be reading its behaviour logs.
Your knowledge base sets the ceiling
Ada can only be as accurate as the content it reads. If you have four return policies from three different eras, two of them contradicting each other, the agent will confidently quote the wrong one half the time. Pruning dead articles, dating the live ones and deleting duplicates is dull work that determines whether you get 65% resolution or 35%. Most failed deployments blame the model. The model usually isn’t the problem.
Where Ada sits against the alternatives
The conversational AI market has consolidated into a handful of serious enterprise players, and the choice usually comes down to workflow complexity and pricing model rather than raw answer quality. Kore.ai builds for enterprises that need heavy orchestration across many internal systems, while Yellow.ai leans into broad channel coverage and voice-first deployments.
A rough way to sort them:
- Ada suits high-volume B2C brands with simple transactional intents: order tracking, returns, password resets, billing questions.
- Kore.ai suits regulated or complex environments where an agent coordinates across a dozen systems.
- Yellow.ai suits companies that want one agent on WhatsApp, voice and in-app chat without building each channel separately.
- Suite-native options from Zendesk and Intercom are the cheapest path if you’re already locked in and your volumes are modest.
Switching costs are real. Migrating thousands of intents from one platform to another is a six-month project, not a weekend.
What Ada costs, and why nobody will tell you
Ada doesn’t publish list pricing. Quotes come from a sales conversation, and the structure has shifted toward outcome-based pricing, where you pay per automated resolution rather than per seat or per conversation. That aligns incentives reasonably well: the vendor gets paid when the AI actually finishes something.
Run the arithmetic before you get excited. Suppose a human-handled contact costs your business $7 all-in and the vendor charges $1.75 per automated resolution. Resolve 80,000 contacts a year and you’re looking at roughly $140,000 in fees against $560,000 in avoided cost. Subtract implementation, integration work, the analyst who maintains the knowledge base, and the helpdesk licence you still pay for. The return is real, just rarely as dramatic as the first slide suggests.
Systems like this keep improving, quietly
Every unresolved conversation is training data. When an agent escalates, the human’s answer gets logged, and good deployments feed those transcripts back in to close the gap. It’s the same loop researchers are chasing in more general settings, where tools like Adaption’s AutoScientist aim to let models improve their own training data. In support, the loop is simpler and more measurable: track which intents fail, fix the underlying content or system connection, and watch the resolution rate climb month over month.
The five numbers to watch in your first 90 days
- Resolution rate by intent, not blended. A healthy overall figure can hide a catastrophe on billing disputes.
- Escalation rate with reason codes. The reasons tell you what to fix next.
- Repeat contact rate within 72 hours. If the same customers return with the same problem, nothing was resolved.
- CSAT split by AI and human. The gap between them matters more than either score alone.
- Cost per resolved contact. Compare it to your pre-launch baseline, including the human hours you didn’t remove.
Pick one boring intent and get it right
The teams that get value out of Ada don’t launch across 40 intents on day one. They pick “where is my order” because it’s high volume, low risk and backed by data that’s already clean. They run it alongside humans for a month, watch the escalation logs, fix the two articles causing most of the confusion, and only then move to returns. Billing comes last, because that’s where the money and the lawyers live.
Eighteen months in, those teams usually have an agent handling the majority of contacts and a human team doing genuinely difficult work. The ones that skip the boring phase end up with an expensive chatbot, a frustrated support team, and a board deck explaining why the AI project is on pause.