The MSP automation market splits into four camps. Each solves a real problem and leaves the same gap: a way to learn how your desk works, prove it is right, and earn its way into your tools one category at a time. That gap is the product.
Rewst, Bumblebee. Powerful and deterministic. You draw the flow, you write the Jinja, you own the orphaned workflows. They need a dedicated automation owner and cannot read a messy free-text ticket on their own.
Pia. Prebuilt automations inside the PSA: triage, forms with identity verification, auto time logging. Runs someone else's process; custom work is PowerShell; implementations run weeks and often need a near-full-time admin.
NeoAgent, Rallied, Thread, Mizo. Agents in the queue that execute. Approval models vary from tiers to "human review". They ask you to trust a resolution rate measured somewhere else, and most bill per ticket, per credit, or per token.
ConnectWise, Kaseya, HaloPSA. Copilots and agents bundled into the platform. Mostly suggest-only or runbook-encoded, tier-gated, and locked to that one PSA. Good if you never plan to switch. Learning comes from the vendor's corpus, not your desk.
A fair reading of each camp's public material as of September 2026. Every vendor is moving fast, so treat the columns as positions, not permanent facts.
| DigbyAI teammate, overlay | Workflow buildersRewst | Packaged packsPia | AI techniciansNeoAgent · Rallied · Thread | PSA-native AIConnectWise · Kaseya · Halo | |
|---|---|---|---|---|---|
| Learns how your desk closes tickets | Memory per client, MSP-wide rules, runbooks drafted from your resolutions, all correctable in plain English | You encode the process yourself | Configuration of the vendor's packs | Early "agent memory" in one product; natural-language workflows in another | Trained on the vendor's corpus |
| Shadow mode with a scored track record | Shadow decision on every ticket, scored on close, calibrated per client and category | Not applicable | No | Shadow pilots offered; no per-category scoring published | Confidence thresholds on triage only |
| Approval model | Every change to a machine or an account is a card with a measured preview; ticket filing and internal notes are switches you set. Autonomy is a dial per client and category, earned on evidence | Whatever you build into the flow | Pack rules and identity verification | Propose / approve / autonomous per workflow, or tiers | Human review, or runbook-encoded |
| Diagnoses before it acts | 34 read-only scripts chosen by symptoms, peer comparison, internet-facing checks, evidence cited line by line | Runs the flow you built | Fixed pack steps | Varies; mostly identity and known fixes | Summaries and suggestions |
| Narrated audit trail | Timestamped note per action with script, device, preview numbers, approver, verify result; time entry posted | Flow logs | Ticket notes | Ticket notes, varies | AI-created flags on records |
| Runs inside your existing stack | Browser badge on the PSA page; a connector per platform (AlgaPSA, HaloPSA, Level, NinjaOne, Hudu, Microsoft 365, Huntress), each in its own process, every read searchable and scoped to the client. Any stack, no rip-and-replace | 80+ integrations, trigger-based | Inside CW, Autotask, Halo | In the queue; some force their own inbox | One PSA only |
| Setup effort | Connect tools, one RMM automation, learning mode in an afternoon, then shadow | Weeks per workflow, dedicated owner | 4 to 8 week implementations | Days to weeks; some need field engineers | Bundled, but tier-gated |
| Pricing model | Flat, per technician, model usage included, BYO key optional | Quote-only platform fee | Quote-only | Per ticket, per credit, or per customer | Bundled; per ticket and per token in places |
● does it · ◐ partly, or depends on what you build · ○ not a goal of that product. Vendor positions summarized from their public features, pricing pages and documentation; nothing here is a claim about quality.
The connectors are commodities. Every PSA and RMM is shipping an MCP server. The governed layer above them, where an agent learns, proves, and earns, is the part that is hard to build and harder to copy.
Digby records what he would have done on every ticket and scores it when the ticket closes. You see match rates per client and per category before he touches anything. Nobody else publishes that number about your own desk.
The trust ladder →Memory with a source on every fact, corrections that outrank anything learned, runbooks drafted from your own closed tickets, and a ledger of every behavior change with a name on it. A track record cannot be copied by a competitor.
Memory and corrections →Every action is narrated to the ticket with the script, the device, the numbers from the preview, the approver, and the verify result. Insurers, auditors, and the technician on the next shift all get the same story.
See the cards →Shadow mode sits beside any of these tools without conflict. Two weeks, a scored report, your call.
Get early access