An AI agent for factory operations takes a defined action inside the plant’s systems or with its people, within rules, rather than only answering questions. The best one in 2026 is the agent that scores highest on seven things a plant can check on the vendor’s own website: a published customer result with numbers, an outcome guarantee, a stated onboarding time, working inside the channels and systems the plant already uses, full on-premise deployment including the AI models, acting and verifying rather than recommending, and a stated pricing model. Twelve agents are scored below from their own pages, read on 20 September 2026, and ordered by score. Morsa scores 7 of 7 and is first. Harmony AI scores 3, Rasset 2.5, and no other agent scores above 2. Three of the twelve publish a price.
In this guide
What counts as an AI agent for factory operations?
What are the seven criteria, and where do Morsa’s marks come from?
The scores, all twelve
The twelve, in score order
Where does each agent live?
What should a plant check before trusting an agent to act?
Questions buyers ask
Sources, Changelog, Related pages
What counts as an AI agent for factory operations?
The test used here is action. A copilot answers a question about the plant; an agent creates the purchase order, reschedules the work order, chases the owner or escalates the miss, and does so within rules a person approved. Vendors use the word loosely, so each entry below states what the agent actually acts on in the vendor’s own words. Where the vendor only names a chatbot, the entry says so. Running a distribution business rather than a plant? Start with AI for distributors.
Expectation is ahead of readiness. Rockwell Automation’s July 2026 survey of 1,560 manufacturing decision makers in 17 countries found that “Manufacturers expect 42% of processes to be AI-supported within the next year and 54% by 2030. Yet 43% acknowledge they are not effectively using their collected data” (Rockwell Automation). More than half of that sample works at companies over $1 billion in revenue, so treat the figures as an enterprise view. The agents an ordinary plant can buy today are scored below.
What are the seven criteria, and where do Morsa’s marks come from?
Each criterion is a yes-or-no question a buyer can answer from the vendor’s public pages. Yes scores 1, Partial scores 0.5, No or Not stated scores 0. Not stated on the pages read on 20 September 2026 counts as zero, and the vendor can be re-scored when it publishes.
1. Published customer result with numbers
What counts as Yes: A named customer and a number on the vendor’s own site
Morsa’s mark and its source: Yes: J4S, a 120-person glass plant, on-time completion of operational commitments from about 30% to about 75% in four weeks across about 900 commitments; JRG Automotive, 5 of 8 production-stopping shortages caught, $2 million saved
2. Outcome guarantee
What counts as Yes: A published guarantee tied to a plant outcome
Morsa’s mark and its source: Yes: “We’ll catch a shortage in your plant, or you don’t pay” (how Morsa is priced)
3. Onboarding time stated
What counts as Yes: The vendor states how long to go live
Morsa’s mark and its source: Yes: J4S went from first connection to live actions in two days
4. Works inside the channels and systems the plant already uses
What counts as Yes: No new app or login for the people who act; any channel, any system
Morsa’s mark and its source: Yes: WhatsApp, Teams, email, SMS or whatever the plant runs on; any system with an API, database, file export or email trail
5. Full on-premise including the AI models and databases
What counts as Yes: Stated on the vendor’s page
Morsa’s mark and its source: Yes: cloud, private cloud, or fully on-premise with the models and databases inside the plant
6. Acts and verifies
What counts as Yes: Takes the action and closes it on evidence, rather than recommending or alerting
Morsa’s mark and its source: Yes: routes, chases, escalates, and closes work on proof such as a photo, a receipt or a system entry; consequential actions follow the customer’s approval rules
7. Pricing model stated
What counts as Yes: The vendor says how it charges, a price or the basis
Morsa’s mark and its source: Yes: value-based, a share of the value a pilot proves, with the guarantee (how Morsa is priced)
The scores, all twelve
1. Morsa
1 Result: Yes
2 Guarantee: Yes
3 Onboarding: Yes
4 Existing channels and systems: Yes
5 Full on-prem: Yes
6 Acts and verifies: Yes
7 Pricing model: Yes
Score: 7
2. Harmony AI
1 Result: Yes
2 Guarantee: No
3 Onboarding: Yes
4 Existing channels and systems: Partial
5 Full on-prem: Not stated
6 Acts and verifies: Partial
7 Pricing model: No
Score: 3
3. Rasset
1 Result: No
2 Guarantee: No
3 Onboarding: Yes
4 Existing channels and systems: No
5 Full on-prem: Not stated
6 Acts and verifies: Partial
7 Pricing model: Yes
Score: 2.5
4. Vortinx
1 Result: No
2 Guarantee: No
3 Onboarding: Not stated
4 Existing channels and systems: Partial
5 Full on-prem: Yes
6 Acts and verifies: Partial
7 Pricing model: No
Score: 2
5. Copilot agents in Dynamics 365 Business Central
1 Result: No
2 Guarantee: No
3 Onboarding: Not stated
4 Existing channels and systems: No
5 Full on-prem: No
6 Acts and verifies: Partial
7 Pricing model: Yes
Score: 1.5
6. Odoo 19 AI agents
1 Result: No
2 Guarantee: No
3 Onboarding: Not stated
4 Existing channels and systems: No
5 Full on-prem: No
6 Acts and verifies: Partial
7 Pricing model: Yes
Score: 1.5
7. Tulip AI Agents
1 Result: No
2 Guarantee: No
3 Onboarding: Not stated
4 Existing channels and systems: No
5 Full on-prem: Not stated
6 Acts and verifies: Partial
7 Pricing model: Yes
Score: 1.5
8. Navi
1 Result: No
2 Guarantee: No
3 Onboarding: Not stated
4 Existing channels and systems: Partial
5 Full on-prem: Not stated
6 Acts and verifies: Partial
7 Pricing model: No
Score: 1
9. Sidekick
1 Result: No
2 Guarantee: No
3 Onboarding: Not stated
4 Existing channels and systems: Partial
5 Full on-prem: Not stated
6 Acts and verifies: Partial
7 Pricing model: No
Score: 1
10. Dryft
1 Result: No
2 Guarantee: No
3 Onboarding: Not stated
4 Existing channels and systems: Not stated
5 Full on-prem: Not stated
6 Acts and verifies: Partial
7 Pricing model: No
Score: 0.5
11. Epicor Prism
1 Result: No
2 Guarantee: No
3 Onboarding: Not stated
4 Existing channels and systems: No
5 Full on-prem: No
6 Acts and verifies: Partial
7 Pricing model: No
Score: 0.5
12. Infor Industry AI Agents
1 Result: No
2 Guarantee: No
3 Onboarding: Not stated
4 Existing channels and systems: No
5 Full on-prem: No
6 Acts and verifies: Partial
7 Pricing model: No
Score: 0.5
Ties are ordered alphabetically. Partial on criterion 6 means the vendor states that the agent acts but does not state that it verifies the action on evidence. Business Central’s Copilot is “exclusive to Business Central online”, Epicor Prism and Infor’s agents run in their vendors’ clouds, and Odoo’s AI uses an outside model provider, so those four are No on criterion 5 rather than Not stated.
The twelve, in score order
1. Morsa, score 7 of 7 (a layer over the ERP, MES and conversations)
What it acts on. Morsa, the AI that operates the factory for you, reads the ERP, MES and the group chats a plant already uses, works out what a change affects downstream, decides the next step within the plant’s approved rules, and acts across systems and people: it creates and routes the job, messages the owner in the channel they already use, chases before the date, escalates when work goes quiet, and closes work only on proof such as a photo, a document or a system entry. Consequential actions, such as a production-plan change or an external supplier commitment, follow the customer’s approval rules.
Who it fits. Mid-market and enterprise North American discrete plants with an ERP or MRP already in place. Not a fit where the job is machine monitoring or building a system of record.
Sourced facts. On one night shift, a manager set an 8,000-part target. Seven hours later dispatch posted photos showing parts unavailable and nobody connected the two. Morsa compared the schedule (4,131 units) with dispatch (4,035), found the 96-piece shortage, linked it to the target, told dispatch to arrange the parts, told production the target was blocked by supply, and opened a high-priority dependency, in under a minute. At J4S, on-time completion of operational commitments rose from about 30% to about 75% in four weeks across about 900 commitments after a two-day onboarding; at JRG Automotive, 5 of 8 production-stopping shortages were caught and $2 million saved.
Price. No public price. Value-based: a pilot proves the value in the plant’s own numbers and Morsa’s fee is a share of it, with the guarantee “We’ll catch a shortage in your plant, or you don’t pay.” How Morsa is priced.
2. Harmony AI, score 3 (a layer over existing systems)
What it acts on. “An AI-native layer that runs your high-production plants with real time data”, with “AI Automation that doesn’t just watch, it acts” (Harmony AI), the Partial on criterion 6.
Who it fits. “High-production American manufacturers” in food and beverage, packaging and printing, building products, consumer products, plastics, paper and distribution, per Harmony’s own list.
One sourced fact. “Customers include Mossberg, MoonPie, and CLS,” with the case study headline “How Mossberg saves 120+ labor hours a week with Harmony”, the Yes on criterion 1, and “A typical pilot runs four to six weeks, with forward-deployed engineers on-site”, the Yes on criterion 3.
Price. No public price; the pricing URL returned 404.
3. Rasset, score 2.5 (its own operating layer)
What it acts on. “Rasset connects every order, material, supplier, worker, machine, payment and decision into a living system” (Rasset), with a “Rasset operations copilot” that is “Connected to live operating records”, the Partial on criterion 6.
Who it fits. “For $10M+ manufacturers, Rasset can become the modern operational layer across facilities.” Plans are labeled “USA FACTORY PLANS”. “Live in days, not quarters” is how Rasset states onboarding, the Yes on criterion 3.
One sourced fact. The only startup in this group with a full price card: “Plans start at $199/factory/month,” rising through Growth at $399, Professional at $799, Scale at $1,499 and $2,999, to Enterprise at “$5,000+”, all per factory per month on annual terms (Rasset pricing).
Price. Published, per factory, from $199 per month.
4. Vortinx, score 2 (a layer, on-premise possible)
What it acts on. “Deploy software agents that process purchase orders, calculate stock safety, log CAPA defects, and alert shift supervisors in real time” (Vortinx). Seven agents are named: Procurement, Inventory, Quality, Maintenance, Production, Reporting and Document.
Who it fits. Automotive, electronics, CPG, pharmaceuticals, chemicals, heavy engineering and aerospace, per its demo form. Its footer places its headquarters in Gurugram, India.
One sourced fact. The strongest deployment statement in this group, and the Yes on criterion 5: “Vortinx can be deployed entirely on-premise without ever touching the public internet,” with a choice of “our secure AWS/Azure Cloud, your private VPC, or bare-metal edge servers on the factory floor.” “Industrial Bridge … sits on top of what you already run” is the Partial on criterion 4.
Price. No public price; the pricing URL redirects to a demo form.
5. Copilot agents in Microsoft Dynamics 365 Business Central, score 1.5 (inside the ERP)
What it acts on. Microsoft’s documentation, dated 3 May 2026, lists a “Sales Order Agent” and a “Payables Agent” alongside Copilot features such as “Chat with Copilot” and “Autofill with Copilot” (Microsoft Learn). These are order-desk and accounts-payable agents; Microsoft’s separate “Factory Operations Agent in Azure AI” is a release-plan item for Microsoft Cloud for Manufacturing, and its own page warns that “Some of the functionality described in this release plan has not been released” (Microsoft release plan).
Who it fits. Business Central customers on the online edition. Microsoft: “With Copilot in Business Central, small and medium-sized organizations empower their workers through intuitive AI tools.”
One sourced fact. “Copilot is exclusive to Business Central online. Therefore, it isn’t available for other deployment types, such as on-premises or private cloud,” the No on criterion 5.
Price. Included: “Copilot in Business Central is included with your Business Central license at no extra cost. However, fair-use policies, quotas, or pricing might be introduced later,” the Yes on criterion 7. Business Central Premium itself lists at $110 per user per month, paid yearly (verified 19 September 2026 on MES vs ERP).
6. Odoo 19 AI agents, score 1.5 (inside the ERP)
What it acts on. Odoo 19, released September 2025, added “Agents: Chat with AI agents that can learn from your documents and perform actions” and “AI agent database queries: Internal users can now chat with an AI agent to query their own database records” (Odoo 19 release notes). The documented actions are inside Odoo: filling fields, server actions, drafting text, generating leads from live chat.
Who it fits. Small and mid-size plants already on Odoo, which has “MRP + MES + PLM + Quality + Shop Floor + Maintenance” on one platform (Odoo Manufacturing).
One sourced fact. “Use your own Gemini account as your Odoo AI provider,” and “ChatGPT 5.0 is now available,” which is why criterion 5 is No: the models are an outside provider’s.
Price.Published: Standard $24.90 billed yearly or $31.10 billed monthly, Custom $49.00 yearly or $61.10 monthly, per user per month for the first 12 months; list $31.10 and $38.90, $61.00 and $76.20, read 24 September 2026 (Odoo pricing, read 21 September 2026). Odoo notes the discount “is valid for 12 months, for initial users ordered.” For what Odoo’s API and plans mean for an outside agent, see Morsa for plants running Odoo.
7. Tulip AI Agents, score 1.5 (on an operations platform)
What it acts on. “Put AI to work in your operations with fully configurable AI Agents that gather information, make decisions, and optimize performance across operations” (Tulip AI), the Partial on criterion 6. The agents act inside Tulip apps and the machines and systems Tulip connects to.
Who it fits. Plants that have built, or will build, their operator apps on Tulip.
One sourced fact. Tulip claims its AI can “Transform SOPs, checklists, and work instructions into interactive apps and cut development time by 80%.”
Price. Published: “$100/mo per interface billed annually” (Essentials) and “$250/mo per interface billed annually” (Professional), “10 interface minimum” (Tulip pricing).
8. Navi, score 1 (a layer over the ERP)
What it acts on. “Navi connects to your ERP, learns your factory, and handles the MRP and scheduling in between, all powered by AI agents and mathematical optimization” (Navi). “Navi’s agent workflow turns each sales order into work orders, handling the routine decisions automatically while keeping a person in the loop for judgment calls,” the Partial on criterion 6.
Who it fits. Plants whose bottleneck is planning and scheduling rather than execution. “Navi plugs into your ERP by API or database, no rip-and-replace,” the Partial on criterion 4.
One sourced fact. The statistics on Navi’s page are labeled by Navi as “Sample Statistics” and “Example statistics”, so none is quoted here as a result.
Price. No public price; the pricing URL returned 404.
9. Sidekick, score 1 (over text messaging)
What it acts on. “The AI agent that manages frontline operations over text” (Sidekick). “Workers text Sidekick in iMessage, WhatsApp, Telegram, KakaoTalk, whatever they already use,” the Partial on criterion 4 because systems are not stated, and it “flags drifting equipment, answers from your own docs, opens and routes work orders, and escalates to the right manager the moment it should,” the Partial on criterion 6.
Who it fits. Shops whose general manager runs the day over text. “No app, no login, any phone.”
One sourced fact. “Shops complete over 15% more jobs,” a vendor claim with no named customer or stated sample.
Price. Not readable: the pricing and FAQ pages sit behind a phone-number sign-in.
10. Dryft, score 0.5 (a layer)
What it acts on. “Dryft builds artificial intelligence that mimics, optimizes, and automates complex human decisions in manufacturing operations” (Dryft), the Partial on criterion 6. Its product animation shows an agent named “Hugo AI” with lines such as “We have 3 late items” and “Follow up with suppliers.”
Who it fits. Not stated. The site is a manifesto and careers page; there is no product page to quote.
One sourced fact. “Announcing our $5M round led by General Catalyst.”
Price. No public price; the pricing URL returned 404.
11. Epicor Prism, score 0.5 (inside the ERP)
What it acts on. Epicor’s launch release of 14 January 2025 describes Prism as “a network of vertical AI agents built specifically for the supply chain industries”, initially “available to be integrated with Epicor Kinetic” (Epicor news release). One named action, the Partial on criterion 6: “Epicor Prism can also automate supplier communications to speed purchasing, enabling users to automatically send RFQs to their supplier network and parse and translate quotes.”
Who it fits. Plants on Epicor Kinetic, in the cloud. Epicor has since named a “Prism EDI Agent” and a “Prism Reasoning Agent”.
One sourced fact. Tyler Madsen, Director at Madsen’s Custom Cabinets, in the release: “Epicor Prism will allow us to reduce many mundane tasks so our staff can prioritize more strategic efforts.” A named customer without a number, so criterion 1 is No.
Price. No public price on either Epicor page read.
12. Infor Industry AI Agents, score 0.5 (inside the ERP)
What it acts on. “Each AI agent applies deep industry expertise, complete data, and secure orchestration to automate workflows, analyze information, and orchestrate multi-step actions within your flow of work” (Infor), the Partial on criterion 6. Infor names eight agent families including “Operations Agents” and “Supply Chain Agents”.
Who it fits. Infor CloudSuite customers: the agents run “securely within your Infor CloudSuite ecosystem”.
One sourced fact. Infor’s page invites customers to a digital event “on October 27 at 10 a.m. EST” under the line “From AI that assists to AI that acts”, which is a fair summary of where the whole ERP group is in 2026.
Price. No public price.
Where does each agent live?
Lives in | Agents | What it can reach |
|---|---|---|
A layer over the ERP, MES and conversations | Morsa | Every connected system, machines where wired, and the people in whatever channel they use |
A layer over existing systems | Harmony AI, Vortinx, Navi, Dryft | The ERP and plant data; channels not stated |
Its own operating layer | Rasset | Records moved into Rasset |
Text messaging | Sidekick | People on any phone; systems not stated |
Inside one ERP | Business Central, Odoo, Epicor Prism, Infor | That ERP’s records, for people with a login |
An operations platform | Tulip | Tulip apps and connected machines |
Also in this space and not scored: Augury’s “Reliability Agent” and role-based agents (machine health), Sight Machine’s “Enterprise Agents” and MCP server (plant data), Plataine’s scheduling agents (composites and aerospace), Lumari (direct procurement agents, not plant floor: “Every action from sending emails to updating the ERP can require team approval”), Allie (six named agents, results anonymized on its site), Tenkara (an assistant named “Rosie” and a “Requote agent”). Oracle NetSuite Next could not be read on 20 September 2026 because netsuite.com and the Oracle newsroom URL both refused the fetch, so it is not described here. For wholesale distribution, see the best AI agents for wholesale distributors (2026). For the wider field, including Palantir, C3 AI and the ERP vendors’ own AI, see best AI software for manufacturing in 2026.
What should a plant check before trusting an agent to act?
Five checks, in the order a plant manager will feel them. They are criteria 6, 4, 6, 5 and 7 in the buyer’s words.
What can it change without asking? Only two of the twelve state an approval model on the page read (Navi and Morsa; Lumari does too). Ask for the list of actions that need a person and the list that do not, in writing.
What does it read? An agent inside the ERP sees the ERP. The shortage that matters is often in a supplier email or a photo in a group chat. Ask which sources are connected on day one.
How does it know the job is done? A “done” message is not proof. Ask what closes a task: a system entry, a receipt, a photo, a supervisor’s confirmation. Morsa is the only agent on this page whose public pages state that work closes on evidence.
Where does it run? Only Vortinx and Morsa state an on-premise option in this group. Under ITAR or CMMC that is the first question, not the last.
What is the price? Three of twelve publish one, and Morsa states its model. A vendor that will not name a price or a basis on a first call is asking you to buy a pilot to find out.
Sight Machine’s MCP server and SAP’s planned “Open-source support for MCP server to enable SAP Business One for AI agents’ adoption with partner extensions” (SAP Business One road map, March 2026) point at where this is going: the systems of record will expose themselves to agents, and the agent a plant chooses will be the one that coordinates across them. The full loop an agent should run, from signal to verification, is on AI copilot for manufacturing.
Sources
Rockwell Automation, 93% of Manufacturers Have MES, But Only 23% Have Fully Integrated It, 14 July 2026, https://www.rockwellautomation.com/en-us/company/news/press-releases/93-of-Manufacturers-Have-MES-But-Only-23-Have-Fully-Integrated-It-New-Rockwell-Automation-Report-Finds.html
Microsoft Learn, Copilot in Business Central, ms.date 3 May 2026, https://learn.microsoft.com/en-us/dynamics365/business-central/copilot-overview
Microsoft Learn, Factory Operations Agent in Azure AI, release plan 2025 wave 1, https://learn.microsoft.com/en-us/industry/release-plan/2025wave1/cloud-manufacturing/factory-operations-agent-azure-ai
Epicor, Epicor Prism news release, 14 January 2025, https://www.epicor.com/en-us/newsroom/news-releases/prism/
Infor, Infor Industry AI, https://www.infor.com/platform/enterprise-ai
Odoo, Odoo 19 Release Notes, September 2025, https://www.odoo.com/odoo-19-release-notes; Odoo Manufacturing, https://www.odoo.com/app/manufacturing; Odoo pricing, https://www.odoo.com/pricing
Tulip, Tulip AI, https://tulip.co/platform/tulip-ai/; pricing, https://tulip.co/pricing/
Harmony AI, https://tryharmony.ai/
Dryft, https://dryft.ai/
Navi, https://naviops.ai/
Vortinx, https://vortinx.com/
Rasset, https://rasset.ai/ and https://rasset.ai/pricing
Sidekick, https://textsidekick.com/
SAP, SAP Business One road map, March 2026 (PDF hosted by an SAP partner), https://www.business-one-beratung.de/fileadmin/user_upload/SAP_Business_One_Road_Map-Maerz26.pdf
Lumari, https://lumari.ai/; Augury, https://www.augury.com/machine-health/; Sight Machine, https://sightmachine.com/; Plataine, https://www.plataine.com/; Allie, https://allie-ai.com/; Tenkara, https://tenkara.ai/
Morsa, J4S customer story, /customers/j4s; how Morsa is priced, /pricing
All pages fetched on 20 September 2026 and quoted verbatim. Business Central Premium pricing was verified on 19 September 2026 for the sibling page and is labeled as such. Every score in the table is derived only from the sentences quoted on this page.
Changelog
24 September 2026: Odoo prices re-read. Billed yearly, Standard is $24.90 and Custom $49.00 per user per month for the first 12 months (lists $31.10 and $61.00); billed monthly they are $31.10 and $61.10 (lists $38.90 and $76.20). Every current Odoo price on this page now names its billing term.
21 September 2026: Odoo price corrected. The ladder is Standard $31.10 and Custom $61.00 per user per month on a yearly term, list $38.90 and $76.20, re-read on odoo.com/pricing on 21 September 2026; the earlier $24.90 and $49.00 figures were Odoo’s previous ladder.
21 September 2026: second draft. Added the seven-criteria scoring table with Morsa’s marks sourced, ordered the twelve by score, and replaced the three location groups with a “where does each agent live” table. Vendor descriptions unchanged.
21 September 2026: first draft. Twelve agents in three groups; NetSuite Next left out because its pages refused the fetch.
