What does a worker actually do?
A worker is a pre-trained AI specialist with chat, tool access, and memory. You give them tasks in plain English; they plan, call tools in your real apps (Salesforce, GitHub, Confluence, Jira, Snowflake, Power BI, Slack, etc.), pause for your call on anything risky, and reply with a complete audit trail of every step.
How do workers hand off to each other?
When you ask Pat to coordinate something, Pat decomposes the goal into a plan and assigns each step to the right teammate. Avery does Salesforce work; Cole writes code; Devin queries the warehouse; the human reviews when needed. Each handoff + tool call lands in the shared activity log. The dispatcher fires independent subtasks in parallel and waits on dependencies before continuing.
Can I connect Dev, UAT, and Production at the same time?
Yes — every multi-instance provider (Salesforce, Snowflake, BigQuery, Databricks, Postgres, MySQL, Redshift, Power BI) supports an arbitrary number of connected instances per worker, labeled with whatever convention you use (production, uat, sandbox-qa, dev). When the user says “update the Account in uat”, the worker routes the call to the right org. Default routing is “production”.
Can I promote changes from one env to another?
For Salesforce, yes. Every write Avery makes is logged with the source instance + record IDs. Ask her to “replay yesterday’s uat changes into production” and she walks the log in order, translating source IDs to target IDs via the map built up from prior creates. Dry-run first surfaces the plan; the real run applies it and records per-row pass/fail.
Whose credentials do they use?
Yours, per worker, per tool. Each worker has their own OAuth connection per integration (and per instance for multi-instance providers). Their permissions match the human user who clicked Connect — no shared service accounts, no permission sprawl. When apsera widens a required scope, every existing customer gets a workspace banner asking them to reconnect — never silent breakage.
How do admins and teammates share a workspace?
Two-tier role model. Owners + admins see every persona on /apps and can hire + assign personas to specific teammates. Members see only their assigned persona on /profile and can connect apps for it, but cannot hire or see other teammates’ personas. Same UI; different scope.
What about my company documents and SOPs?
Drop them in. Each worker has their own knowledge base — PDFs, DOCX, Markdown, HTML, or URLs. We chunk + embed via Cloudflare Workers AI (768-dim BGE), store in pgvector, and ground every reply via RAG with [Doc: filename] citations. Original binaries in Cloudflare R2.
When do I have to step in?
Every action with side effects (writing to Salesforce, opening a PR, sending an email, charging a card, replaying changes across envs) pauses for your approval. The worker explains what they’d do, lists alternatives, and resumes the moment you decide. Read-only work runs without interruption. Approvals are namespace-scoped, so the first approval per provider covers every write in that task — three SF updates = one prompt.
What if a worker hallucinates or loops?
Two guards. The hallucination guardwatches for help-seeking phrases in end-of-turn summaries (“I’ll need you to…”, “cannot proceed…”) and converts them into approval prompts so phantom “done” tasks don’t look complete when they aren’t. The circuit breaker trips after the worker hits the same tool / input failure 3× and pauses for your input instead of looping forever.
How does billing work?
Three modes, all auto-charge via Razorpay. Subscription: pick a tier (Starter $49, Pro $149, Business $399 per worker/mo) and Razorpay auto-debits monthly × active worker count. PAYG: authorise a card once ($0.50 verification), monthly cron sweeps based on token usage. One-time: pay each charge manually. Invoices stored locally with downloadable PDFs. Admins see MRR + invoices + saved cards across all customers in /admin/billing.
Can I bring my own AI / BI license?
For tools that need vendor-specific licensing (Power BI Pro / PPU / Fabric, Salesforce user licenses, Snowflake compute, etc.), you connect your own. apsera.ai integrates the tool — you pay the vendor directly. The pricing tiers cover apsera’s side (workers + LLM tokens + storage); vendor licenses are pass-through.
Is it safe to give an AI access to production?
You decide. Approvals are required by default for every write; the worker can’t bypass them. Read-only flows run on autopilot. Everything is logged with diff + undo for SF updates. Permissions flow through the user’s OAuth grant — the worker can never exceed what the human connected. For Salesforce specifically: the day-1 profile scan also captures every custom field + picklist so Avery never hallucinates a stale value into a write.