Asynchronous-to-Synchronous Channel Conversion

Asynchronous-to-synchronous channel conversion is the replacement of a service channel in which a customer and an agent exchange messages over an open, persistent conversation with one in which they meet in a bounded session that ends when either party stops responding. The conversion is usually undertaken for operational reasons: session-based channels are easier to route, measure and staff on tools built for voice, and they close conversations that would otherwise stay open indefinitely. The planning consequence is less often recognized. The two channels are planned with different models, built on different units of demand, with different meanings for waiting, ending, handle time and capacity. Neither model is a variant of the other, and the history of the first is not a baseline for the second. An operation that converts the channel and carries its forecasts, targets and staffing ratios across will have changed the channel on the customer while continuing to plan as if it had not.

This page states the two models side by side, shows why they cannot be compared on a shared unit, identifies the one layer of history that does transfer, and describes the rebaselining an operation owes the new channel. The mechanics of each model are held on their own pages: Demand Forecasting for Digital and Async Channels for the asynchronous side and the concurrency parameter, Multi-Channel and Blended Operations for capacity per channel and the shape of an asynchronous service level, and Digital Messaging for the interaction models themselves. The wider question of whether customers follow a channel shift at all belongs to Contact Deflection and Channel Shift Modeling; this page concerns the narrower case in which the customer does not move, because the channel moved around them.

The condition

The case is easiest to see in a service where the work is naturally spread across a day. A traveler arranging a trip through a persistent messaging channel opens a conversation in the morning, receives options, steps away, returns at midday with a decision, asks a follow-up in the afternoon, and closes the matter in the evening. To the customer it is one conversation. To the operation it is one open item of work, touched several times, with an agent returning to it between other conversations. Nobody waited in a queue. Nobody abandoned. The conversation simply stayed open until it was finished.

The same traveler on a session-based channel with a ten-minute timeout has a different experience, and the operation records a different set of events. The morning exchange is a session. Stepping away ends it. Returning at midday begins a second session, possibly with a different agent, who reads or does not read the first. The afternoon question is a third. The operation now records three contacts, three handle times, some number of sessions that timed out with the customer intending to return, and, if the customer gave up between sessions, an abandonment that no report will label as one. The customer did not choose a new channel. The customer's habit was formed in the old one, and channel habit is durable: prior use of a channel raises the likelihood of using it again. The new channel therefore receives customers behaving as the old channel taught them to behave, on a channel that punishes that behavior with a timeout.

The two planning models

The models differ at every layer, and the differences are not of degree.

Layer	Asynchronous conversation channel	Synchronous session channel
Unit of demand	A conversation, opened once and touched many times	A session, offered once and either handled or lost
Arrival process	Messages arrive clustered around conversations already open; the next message depends on the last exchange. Forecast the stock of open conversations and the rate of touches	Sessions arrive by interval and are forecast as arrivals; a returning customer is a new arrival
What waiting means	Time until the next response inside an open conversation; the customer is usually doing something else	Time in a queue before the session begins, and time between replies inside the session; the customer is present and counting
What ending means	The customer returns later, or does not; a lapse is a retrial in waiting, and retrials are the measurable unit of outcome in message-based service	The session times out or the customer leaves; the lapse is an abandonment, including abandonment during service, a failure mode with no counterpart in voice queueing and, by this page's reading, none in an open conversation either
Capacity unit	Open conversations an agent can carry, bounded by response-time tolerance; work in progress	Concurrent sessions an agent can hold, bounded by abandonment during service and by the quality cost of multitasking
Handle time	Agent work time across the conversation's life, with re-orientation on each return; elapsed time is days and means little	Session time, co-produced: the customer's own reply speed is part of it, and it slows when the customer has been kept waiting
Service measure	Response time within tolerance, and the rate at which customers come back to an unresolved conversation	Answer time, and abandonment, with the corrected abandonment including customers who left silently
Staffing method	Flow: touches per interval against agent work time, with the backlog carried; a deferred-work model	Not Erlang, although it is usually planned as if it were. The literature on chat staffing replaces the one-agent-one-contact assumption with processor-sharing models and routing that minimizes the proportion abandoned
What occupancy means	Share of agent time on touches; idle time between touches is the normal state of a deferred channel	Ill-defined under concurrency; an agent holding three sessions is neither fully occupied nor a third idle

Three of the rows carry most of the weight. The unit of demand is the first: a conversation and a session are different objects, and a count of one is not a count of the other. The meaning of ending is the second: the customer who steps away from an open conversation has done nothing, while the customer who steps away from a session has abandoned it, and the operation's reports will treat the two as different customers doing different things when it is one customer doing the same thing. The staffing method is the third, and the least visible. The synchronous channel invites the voice planner's toolkit, and the toolkit's core assumption, one agent occupied by one contact, does not hold for it. The peer-reviewed literature proposes no adjusted Erlang for chat; it abandons Erlang for it. A planner who applies a concurrency multiplier to an Erlang result has not adjusted a valid model but borrowed an invalid one.

Why the models cannot be compared

A comparison needs a shared unit, and there is none. The conversions that seem to supply one turn out to be properties of the channel change rather than of the demand.

The number of sessions a former conversation becomes depends on the timeout, on the customer's habit, and on the time of day the customer works; it is not a fact about the customer's need. A handle time on the asynchronous side is agent work time spread across returns; on the synchronous side it is session time with the customer's pace inside it, and the two are not the same quantity measured in different places but different quantities that happen to share a name. No settled definition exists for which of the two a chat handle time should carry, and a figure carried across the conversion without its definition cannot be interpreted. A retrial on the asynchronous side, a customer returning to an open conversation, is invisible in a session-based report except as a new contact, and an abandonment on the synchronous side has no asynchronous counterpart to be compared against. Even the population is not held constant: the customers of the new channel are the customers of the old one, behaving as the old one taught them, for as long as habit persists and no longer.

It follows that the old channel's volume, handle time, concurrency, abandonment, patience, service level, occupancy and forecast model do not transfer. They are not wrong by a factor to be estimated; they are measurements of a different object. An operation that carries them across is not making an error of degree that a variance report will catch. It is planning one channel on the history of another and will be surprised by the difference in a form its reports were not built to show.

What does transfer

One layer of the old channel's history survives the conversion, and it is the layer beneath the channel. Customers still want the same things, at roughly the same times, in roughly the same mix. The intent, the time the customer first reaches out, the seasonality of the need and the population reaching out are properties of the demand, not of the channel, and the old channel's record of them is the best evidence available for the new one. The forecast that transfers is therefore a forecast of intents by interval of first contact, with the channel layer stripped off. Everything above that layer, how many sessions an intent becomes, how long each takes, how many are lost, is a property of the new channel and is unobserved until the new channel has run.

The agents' knowledge transfers, with one caution. Knowledge of the products, the customers and the systems carries across. Whether an agent's pace and quality carry across from a deferred channel to a present-customer one is not established; no study measures that transition, and the assumption that it is costless rests on capacity models that assume it rather than evidence that shows it. Task Switching Costs in Multichannel Operations describes the cost of moving between channel modes within a day; the conversion asks agents to move permanently, and the ramp is an open question rather than a known parameter.

Rebaselining

The practice that follows is to treat the converted channel as a new channel with a history that begins on the day of conversion, and to plan the first period on stated assumptions rather than on carried history.

• Declare the baseline reset. The new channel's targets, forecast model and staffing ratios are marked open on the day of conversion, and every figure carried from the old channel is labeled as carried and as inferred, in the sense of the graded register on Data Synthesis Before Decision. A carried figure is a hypothesis about the new channel, not a measurement of it.
• Forecast intents, not sessions. The demand forecast for the measurement period is the old channel's intents by interval of first contact. Sessions per intent is the first parameter the new channel has to reveal, and it is estimated, not assumed.
• Staff the measurement period on a stated model. Provisional staffing is set on a flow model of intents and an assumed sessions-per-intent, with concurrency deliberately below the tool's cap, because abandonment during service is the cost of running at the cap. The assumption is written down so that it can be found wrong.
• Instrument what the reports will not show. Four measures the session-based channel does not produce by default, and that decide whether the conversion worked: the corrected abandonment, including customers who left silently; repeat sessions on the same intent within a window, which are the old channel's retrials made visible; handle time on both definitions, elapsed and agent work; and timeouts on sessions the customer subsequently resumed, which count customers the channel closed against their intent.
• Set targets after the period, not before. Service level, abandonment and handle time targets for the new channel are set from its own first period, with the measurement period long enough for customer habit to begin adapting and short enough that the operation is not running blind. Targets inherited from the old channel, or from voice, are the failure mode the practice exists to prevent.
• Govern it as an initiative. The conversion is an initiative with a volume and service impact, and the governance Contact Deflection and Channel Shift Modeling prescribes for deflection applies: a documented method, review by the forecasting function before the plan changes, and variance tracked against the stated assumptions afterward.

The failure mode

The pattern in this section is a prediction from the mechanisms above, not an observed result; no study has measured a conversion. If the mechanisms hold, an operation that converts the channel and keeps its plan would understaff the new channel, and the shortfall would not appear where the reports look. Sessions would time out and customers return, so recorded volume would rise while the intent count had not moved, and the operation would read growth. Handle time per session would be recorded as shorter than handle time per conversation, because a session is a fragment of one, so productivity would appear to improve; no matched-intent comparison of the two exists to say by how much. Customers who left between sessions would leave silently, so abandonment would appear low. The service level, measured on the sessions that were answered, would hold. Each inherited instrument would report a healthy channel while measuring a different object from the one the customer was experiencing. The finding would surface later, as complaints, as repeat contacts on other channels, or as a satisfaction measure that moved for reasons no operational report explains. Instrument Effects During Measurement Rollout describes the case of an instrument changing beneath a stable operation; the conversion is its mirror, a stable instrument reporting on an object the operation changed.

The opposite failure is smaller but real. An operation that rebaselines on caution alone, without instrumenting, will overstaff the session channel against the sessions it now sees, including the sessions that are fragments of one intent, and will conclude that the new channel is more expensive than the old. It may be, but the count of sessions is not the evidence for it.

What the evidence supports and what it does not

The argument on this page is a chain, and each link rests on a measured result while the chain as a whole does not. That message-based conversations arrive as clustered, history-dependent exchanges rather than as independent arrivals is established on a large operational dataset. That in the text channels studied, customers waited long by voice standards and left silently in large proportion, so that recorded abandonment understates the true rate and agents carry conversations that have already ended, is shown with a stated method on real contact-center data, though not yet in a peer-reviewed venue. That abandonment during service exists in session-based chat and holds optimal concurrency below the technical cap is peer-reviewed. That multitasking across sessions lengthens in-service delay and lowers resolution, and that a waiting customer slows their own replies, are field results in peer-reviewed journals. That the standard voice staffing model does not extend to chat is the settled position of the operations research literature on the subject.

What no study measures is the conversion itself. The literature treats asynchronous and synchronous channels as coexisting choices a customer or an operation makes, and contains no before-and-after account of an asynchronous channel being converted to a synchronous one: no measured fragmentation of conversations into sessions, no measured repeat-session rate, no measured change in satisfaction or abandonment attributable to the conversion. Nor does any study compare handle time across the two channels for matched intents, or measure the ramp of agents moved between them. The silence has a use. An operation performing the conversion is producing evidence that does not otherwise exist, and the instrumentation in the rebaselining section is what would turn its experience into a finding.

Limitations

The page concerns a conversion imposed on an existing customer population, and its argument weakens as that population turns over or adapts; a session channel offered to customers who never used the conversation channel has no inherited habit to contend with, and its planning problem is the ordinary one of a new channel. The argument also assumes the operation wants the converted channel to serve the same intents; where the conversion is accompanied by a deliberate narrowing of what the channel is for, the intent forecast does not transfer either, and the rebaselining starts further back. The four instruments proposed are the minimum that would make the conversion observable, not a full measurement design, and the measurement period's length is a judgment the page does not make for the reader. Finally, the two models are stated as they appear in the literature and in practice; an operation whose asynchronous channel was already run on session-like rules, with enforced response windows and closures, has less distance to cover than the page assumes.

Maturity Model Position

In the WFM Labs Maturity Model™, the conversion is most dangerous at Level 2, because that is the level at which an operation has enough reporting to be reassured by it, and a definitional discipline aimed at holding definitions stable across systems rather than at detecting that the object beneath a stable definition has changed. The rebaselining practice is Level 2 work done by hand: a declared reset, a stated model, four added measures. At Level 3 the real-time layer can enforce concurrency below the cap and surface silent abandonment as it occurs, and the intent-level forecast that transfers across the conversion is the kind of channel-independent demand object a Level 3 forecasting function should already maintain. Level 4 plans the measurement period as a distribution over sessions per intent rather than a point assumption, and Level 5 would treat the channel design itself, including the timeout, as a parameter the operation sets on evidence rather than inherits from the tool.

See Also

• Demand Forecasting for Digital and Async Channels — the asynchronous forecasting mechanics and the concurrency parameter, both assumed here
• Multi-Channel and Blended Operations — capacity per channel, why concurrency is not a simple multiplier, and the asynchronous service-level shape
• Digital Messaging — the synchronous and asynchronous interaction models as channels
• Contact Deflection and Channel Shift Modeling — the wider question of whether customers follow a channel shift, and the governance a channel initiative owes the plan
• Abandonment Rate Modeling and Patience Distributions — patience by channel and re-contact from abandonment, the synchronous side's ending
• Blending and Deferred Workload — how deferrable work is absorbed into inbound idle capacity, the treatment an asynchronous backlog receives in a blended operation
• Task Switching Costs in Multichannel Operations — the cost of moving agents between channel modes
• Instrument Effects During Measurement Rollout — the general case of a measure changing under the operation
• Data Synthesis Before Decision — the graded register in which carried figures are marked inferred
• Human Gates and Number Grades — the carried-assumption rule generalized: nothing carried across a channel, platform or cohort change is measured in the new state