Einstein Bot to Agentforce: convert or keep the menu?

Convert an Einstein Bot to Agentforce only where customers type questions the bot never understood, and keep short, rule-bound paths menu-driven. Salesforce’s Upgrade Agent, announced at Dreamforce 2026, builds an Agentforce Service Agent from your bot in minutes while the bot keeps running, but it is a Labs tool and its output is a draft. The bigger changes sit around the tool: cost moves from a licence allowance to Flex Credits per action, behaviour moves from scripted dialogs to a model that decides, and testing moves from clicking through paths to pass rates.

Do you have to replace your Einstein Bot now?

I found no deadline. On the Salesforce pages I checked by 27 October 2026, Einstein Bots carry no retirement date. The pressure came from the channel instead: Salesforce’s developer documentation says legacy Chat was “scheduled for retirement on February 14, 2026”1, with Messaging for In-App and Web as the successor. Salesforce’s own implementation guide for customer-facing agents2 deploys its example agent on an Enhanced Web Chat channel inside an Experience Cloud site, which is the setup many manufacturer portals already use.

The tool itself is item 20 in Salesforce’s list of 21 Dreamforce 2026 announcements3: “Click Upgrade Agent on Agentforce Labs to convert an existing Einstein Bot into an Agentforce Service Agent.” The post adds that it “builds the new agent in minutes, so your current bot keeps running while you review and adjust the upgraded version.” It names no GA or beta status. On its Agentforce Labs page4 Salesforce writes: “Everything in Labs is experimental and pre-release.” I would run it in a sandbox and treat the result as a first draft. Which parts of a bot carry over, such as dialogs, variables or Apex actions, is not described on any page I could open, so check it in your org before you estimate effort.

The script becomes a model that decides for itself

An Einstein Bot follows a script. Salesforce described dialogs as “the basic ‘unit’ of conversation”5, intents as example phrases per dialog and rules as “a conditional engine” for runtime decisions, and bots can “present customers with pre-defined options in menus”. When a dealer types something no intent matches, the bot says it did not understand and shows the menu again. Annoying, but visible.

An Agentforce Service Agent works with topics, instructions and actions, and a language model chooses the action and phrases the answer. That handles questions like “The spindle motor on serial 4711 failed again, is the part in stock and can a technician come this week?” It also changes how the system fails. An agent rarely says it does not understand. If the knowledge article is outdated or the delivery date from SAP is a week old, it gives a fluent, wrong answer. In machine building the expensive cases are warranty statements, delivery promises and anything that touches safety.

Two consequences:

• Decisions with consequences move into deterministic logic behind the agent, for example warranty eligibility or goodwill limits. I describe the mechanics in Agent Script rules instead of LLM decisions.
• Knowledge becomes a production system. A bot showed the one article a dialog pointed to. An agent reads and rephrases whatever grounding it gets, so every manual and article needs an owner and a review date.

The allowance becomes consumption

Einstein Bots were budgeted as an allowance. Salesforce’s Einstein features cheat sheet6, updated April 2024, says “Service Cloud UE licenses come 25 Einstein Bot conversations PUPM” (per user per month), Enterprise Edition gets them only with the Digital Engagement add-on, and extra capacity was “100 conversations for $50 per month”.

Agentforce bills consumption. The Agentforce pricing page7 lists Flex Credits at $500 per 100,000, and Salesforce’s May 2025 announcement8 says “One Agentforce action consumes 20 Flex Credits ($0.10 USD)”. The alternative is $2 per conversation, and the pricing page states that Flex Credits and conversations “will not be supported in the same org”. Prompts and Data 360 grounding come on top of actions; I break them down in Flex Credits per action.

There is a third model. The rate card9 lists Help Agent resolutions at 400 credits, and Salesforce’s June 2026 announcement10 says there is no charge when the customer is handed to a human. Salesforce announced this for Help Agent, now marketed as Casey, which I cover in the Casey article. I found nothing saying it applies to a custom service agent built from your bot.

The Service Cloud pricing page11 lists Max with “2.75M Flex Credits per org per year”. That pool is shared by every agent and every Data 360 job in the org.

Worked example: a dealer portal bot with 4,000 conversations

Assume a machine builder with 300 dealers and service partners on an Experience Cloud portal. All numbers are assumptions for illustration. An Einstein Bot handles 4,000 conversations a month:

Bot path	Share (assumed)	My recommendation
Order status by order number, SAP lookup	30 %	keep in menu
Manuals and certificates by serial number	15 %	keep in menu
Warranty registration of a new machine	10 %	keep in menu
Machine breakdown report	10 %	capture fixed fields, hand to a human
“Something else”, free text	35 %	convert to the agent

Option 1, convert everything. At an average of three actions per conversation, 4,000 conversations need 240,000 credits a month, $1,200 at list price, or 2.88 million credits a year for actions alone. That exceeds the 2.75 million in a Service Cloud Max org before a single prompt, grounding job or test run. At $2 per conversation the same volume costs $8,000 a month.

Option 2, hybrid. The bot keeps the first four paths, 2,600 conversations. The 1,400 free-text conversations go to the agent. Free text tends to need more steps, so assume four actions: 112,000 credits, $560 a month, about 1.34 million a year.

2.88M

credits a year for actions if everything is converted

1.34M

credits a year for actions in the hybrid option

2.75M

Flex Credits per org and year in Service Cloud Max

Cost is the smaller argument for option 2: the free-text bucket is where the bot fails today, while the menu paths already work. A December 2025 post on the Salesforce Developers blog about handing off from Einstein Enhanced Bots to Agentforce agents, written by two Salesforce employees, describes this pattern with Einstein Enhanced Bots: the bot “triages transactions (Log/Status) and routes complex knowledge to an Agentforce Service agent”, using a field on the Messaging Session, an outbound Omni-Channel flow and a fallback queue. The post covers only the handoff from bot to agent, not the way back.

Which flows should stay menu-driven?

My criteria, in this order:

• Identification and verification. Customer number, dealer ID and serial number decide what the conversation may see. That should not depend on how a model reads a sentence, and free-text entry points are where prompt injection in dealer portals starts.
• Data capture with legal weight. Warranty registration, commissioning date, the contact for safety notices. A menu or form captures exactly the mandatory fields.
• Safety-relevant reports. A breakdown with risk to people goes straight to a human queue with serial number and photos, without a summary in between. For the phone channel I cover the same question in the breakdown hotline article.
• Single lookups with one key. Order status by order number is one SAP call. An agent can answer it too, as described in the SAP order status article, but it spends two or more actions on an answer the menu gives in two clicks.
• Promises with liability, such as confirmed delivery dates.

Convert to the agent: troubleshooting questions answered from manuals, mixed questions, dealers writing in their own language and follow-up questions in the same conversation.

If you move everything to Agentforce anyway, the deterministic paths can live inside the agent as Agent Script logic, but each step then runs on the agent’s meter.

How do testing and go-live change?

A bot is tested by walking each dialog path once: same input, same output. An agent needs a test set that runs several times against an agreed pass rate. The implementation guide scores response quality from 0 to 5 and counts 3 or higher as a pass, while topic and action checks are pass or fail. Thresholds and tooling are in testing Agentforce agents before go-live. Test runs in a sandbox cost 16 credits per action according to the rate card.

The bot gives you a head start. The example phrases each intent was trained with, plus the transcripts of the “something else” path, are a ready-made test set of real dealer questions.

This split is an architecture decision worth making before the first upgrade click. The order I would follow:

1. 01

   Review the paths

   Export each bot path with volume and drop-out rate for the last three months.

2. 02

   Mark the paths

   Mark every path as keep, convert or hand to a human, using the criteria above.

3. 03

   Generate the draft

   Run Upgrade Agent in a sandbox and cut the draft down to the converted paths.

4. 04

   Build the tests

   Build tests from intent phrases and transcripts and agree the pass rate with the head of service.

5. 05

   Go live with free text

   4 weeks

   Route only the free-text path to the agent through Omni-Channel with a fallback queue, and watch Digital Wallet.

6. 06

   Pick the next path

   Decide on the next path with those numbers.

Sources11

1. 1“scheduled for retirement on February 14, 2026” (opens in a new tab)developer.salesforce.com
2. 2implementation guide for customer-facing agents (opens in a new tab)resources.docs.salesforce.com
3. 321 Dreamforce 2026 announcements (opens in a new tab)salesforce.com
4. 4Agentforce Labs page (opens in a new tab)salesforce.com
5. 5“the basic ‘unit’ of conversation” (opens in a new tab)developer.salesforce.com
6. 6Einstein features cheat sheet (opens in a new tab)salesforce.com
7. 7Agentforce pricing page (opens in a new tab)salesforce.com
8. 8May 2025 announcement (opens in a new tab)salesforce.com
9. 9rate card (opens in a new tab)salesforce.com
10. 10June 2026 announcement (opens in a new tab)salesforce.com
11. 11Service Cloud pricing page (opens in a new tab)salesforce.com